DEV Community: AI Feed

OpenAI Trains GPT-6 Astra on Ironclad Contract Tasks

AI Feed — Mon, 19 Oct 2026 18:05:14 +0000

OpenAI trained Astra using 11 contract-workflow tasks developed with Ironclad, reporting higher scores but only simulated timing estimates.

OpenAI said on October 6, 2026 that it trained GPT-6 Astra on contract-workflow tasks developed through a research collaboration with Ironclad. Ironclad provided hosted versions of its software, while OpenAI created the synthetic assignments and used reinforcement learning to train the model. The research examines how agents handle complex work in specialized professional applications.

The companies designed 11 tasks drawn from legal, commercial and procurement processes. They included preparing a nondisclosure agreement configuration, setting up approvals for purchasing and adapting reusable contract language to a jurisdiction selected by a requester. OpenAI estimated that each assignment would take an experienced user about 30–40 minutes.

Performance was judged against 8–50 criteria per task. Rather than considering only the final output, the assessment checked whether an agent observed applicable business rules, moved through all required stages and ultimately fulfilled the request. This structure was intended to test completion of a professional workflow, not text generation in isolation.

According to OpenAI, Astra averaged 55.0% on the scoring rubrics, versus 41.6% for GPT-5.6 Sol, a relative increase of 32%. Simulated time per attempt declined by 48%, from 37.0 minutes to 19.2 minutes. A nonpublic model used in Astra’s development reached 63.7%, and OpenAI said it plans to carry those improvements into future models.

OpenAI confined the findings to this 11-task set; they do not represent performance across every Ironclad workflow. The duration figures were calculated using assumed model processing and generation rates, rather than observed savings for customers. OpenAI also noted that an agent can overlook business rules as a complex task unfolds, so human review remains important.

The simulated assignments were based on public contracts from SEC EDGAR, with personal information filtered. OpenAI said the research excluded customer data, OpenAI’s internal agreements, and nonpublic contracts or other customer information held by Ironclad.

The results offer evidence about evaluating an agent across an end-to-end process, but the restricted benchmark and modeled timing prevent the reported acceleration from being treated as a demonstrated improvement in live contract operations. OpenAI’s announcement also does not make either the 11 assignments or the internal model that scored 63.7% available as standalone public products.

Model	Mean score	Estimated time per attempt
GPT-5.6 Sol	41.6%	37.0 minutes
GPT-6 Astra	55.0%	19.2 minutes
Internal Astra development model	63.7%	Not specified

Sources

openai.com

Originally published by AI Feed: OpenAI Trains GPT-6 Astra on Ironclad Contract Tasks.

OpenAI Publishes Mathematics Results From an Internal Model

AI Feed — Mon, 19 Oct 2026 14:33:18 +0000

OpenAI released AI-generated mathematical results, Lean formalizations for many proofs, reasoning summaries and estimates of computing effort.

OpenAI has published a broad collection of mathematical results generated by an internal frontier model, giving mathematicians material to examine, correct and build upon. The October 6 announcement does not identify the model by name or version, and the system remains internal while the company works toward what it calls a responsible release.

The collection is hosted in a GitHub repository with defined procedures for revisions and citations. OpenAI said it consulted the independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study and used the group’s public recommendations to shape the release.

Many proofs are accompanied by Lean formalizations, allowing those particular arguments to be checked computationally. That coverage is incomplete: not every proof currently has a formalization, and OpenAI says it will add more as they become available. Publication in the repository therefore does not mean that every result has already received formal machine verification.

The repository also includes ten summaries of the model’s reasoning, statistics on attempted problems and estimates of computational effort expressed through ChatGPT Pro usage. According to OpenAI, producing an average result required compute roughly equivalent to three hours of ChatGPT Pro thinking. The disclosed reasoning is summarized rather than a complete record of every model process.

For researchers, the combination of conventional mathematical exposition, Lean files and attempt statistics provides different ways to assess the work: readability, computational checkability and the apparent cost of finding a result. Reproduction remains constrained, however, because the model is still internal and only reasoning summaries—not complete process records—have been released.

In an update dated October 15, OpenAI announced a $5 million contribution to a new Simons Foundation initiative supporting mathematicians who examine, understand and extend AI-produced results, including this collection. The foundation will independently manage and award funding for workshops, conferences and special programs. OpenAI also said supported researchers would receive access to models and computing resources.

OpenAI describes the publication method as unfinished. It is exploring community-hosted alternatives that meet the advisory group’s guidelines, while acknowledging that future papers need better citations, mathematical exposition, presentation and overall comprehensibility. The company did not provide a release date for the internal model and said it would continue incorporating community feedback into standards for disclosing major scientific advances.

Sources

openai.com

Originally published by AI Feed: OpenAI Publishes Mathematics Results From an Internal Model.

Radisson Adds Hotel Search and Comparison to ChatGPT

AI Feed — Mon, 19 Oct 2026 11:03:15 +0000

Radisson’s plugin shows hotel locations, prices and amenities in ChatGPT, but travelers are redirected to the company’s website to book.

Travelers can now use a Radisson Hotels plugin to find and compare the group’s properties without leaving a ChatGPT conversation. Radisson Hotel Group developed the integration with Accenture Song, according to OpenAI’s October 7 announcement. Hotel options appear in ChatGPT, while the booking transaction is completed separately on Radisson’s website.

The results can include a map, room pricing and information about property amenities. OpenAI says the team took six weeks to bring the plugin from development to launch, using an agentic development approach. Accenture supplied the MCP server and APIs behind both the hotel-search tool and its connected advertising experience. The companies intend to reuse that technical foundation for later additions.

OpenAI reported an early performance comparison covering July and August 2026. During those two months, the share of plugin visits that became bookings was about 1.5 times the corresponding rate for Radisson’s organic search traffic. The announcement provides no booking totals, sample size or data from a longer period, so the result does not establish whether the difference will persist or how much business it represents.

The rollout also has a paid-marketing component. Radisson began sponsored campaigns in ChatGPT in July and subsequently extended them to more markets in Europe and elsewhere. OpenAI says view-through measurement assigned 54% of recorded checkout and booking events connected with those campaigns to ad exposure. Because this method can credit a viewed ad without a click, the percentage is an attribution measure, not evidence that advertising by itself caused those transactions.

Users who have installed the plugin no longer need to call it manually with “@RadissonHotels.” Following an update, ChatGPT may present it when a request is relevant. That behavior is limited to existing plugin users; the announcement does not say Radisson results will appear automatically for everyone using ChatGPT.

The integration reduces the number of interfaces needed to move from general trip planning to a comparison of specific Radisson properties. Its limits are equally important: checkout remains outside ChatGPT, and the only performance figure supplied is a two-month relative rate reported by OpenAI, the provider of the platform and advertising product. Without absolute volumes or longer-term data, the commercial impact cannot yet be assessed from the announcement alone.

OpenAI presents the deployment as an initial step rather than a finished expansion of hotel booking inside ChatGPT. Radisson’s stated priorities are to encourage wider use of the plugin and introduce more functions within the chat experience. Those objectives describe future work, not features confirmed as available in the current release.

Sources

openai.com

Originally published by AI Feed: Radisson Adds Hotel Search and Comparison to ChatGPT.

How to Build a Resilient AI API Integration: Timeouts, Retries, Queues, and Graceful Degradation

AI Feed — Mon, 19 Oct 2026 09:46:23 +0000

One successful model request proves only that the request worked once. In production, an integration must withstand 429 responses, provider overload, network timeouts, interrupted streams, and traffic spikes without turning a temporary failure into a retry storm. The practical approach below is to classify errors, limit both elapsed time and attempt count, separate workload types, control concurrency, and define degraded modes in advance.

A demo tests a feature; production tests a system

A prototype usually has one user, a short context, and almost no competing requests. In a production product, several factors combine: one user uploads a large document, a scheduler starts a batch job, a chat service maintains dozens of streaming connections, and all clients retry simultaneously after a brief outage.

AI API quotas are often multidimensional. Providers may limit not only request counts but also input or output token volume. For example, the Gemini API rate-limit documentation describes limits based on requests and input tokens, and exceeding any one dimension is enough to trigger an error. Exact values depend on the model, project, usage tier, and available capacity, so a universal quota table would quickly become outdated.

Reliability here is not a promise that errors will never occur. It is the system’s ability to contain the scope of a failure, preserve important functions, and recover without creating additional load.

Classify the error before choosing an action

An HTTP status alone is insufficient. The client should inspect the machine-readable code in the response body, relevant headers, and the provider’s specific rules. The same 429 status can represent a temporary rate restriction, an exhausted daily quota, or a financial limit. An immediate retry is not appropriate in all of these cases.

Situation	Retry?	Delay	Limit	Fallback
Invalid parameters, unsupported model, or schema error	No, not until the request is corrected	Not applicable	0	Validation and an explicit error for the developer or user
API key, permission, or billing problem	No, not until the configuration changes	Not applicable	0	Stop the task and send an operational alert
429 with a temporary restriction and Retry-After	Yes, if the operation still fits within its deadline	No shorter than Retry-After, plus a small amount of jitter	Within the retry budget	Fast degradation or a queue
429 caused by a daily quota or hard spending limit	Not immediately	Until the quota resets or the limit changes	0 rapid retries	Defer the work or report that the service is unavailable
Temporary 5xx response or model overload	Usually	Retry-After or bounded exponential backoff with jitter	A small number of attempts within the deadline	Another tested configuration or a reduced response
Timeout before a result is received	Only if repeating the operation is safe	Backoff followed by a check of the remaining time	Within the overall budget	Queue the task, show a “processing” status, or return an explicit error
Interruption after part of a stream has been received	Do not replay automatically	—	0 hidden retries	Show the partial result and offer an intentional continuation
Content block or deterministic rejection	No for the same input	—	0	Safely change the workflow or explain the refusal

This matrix must be adapted to the specific operation. Repeating draft generation usually does not change external state. Repeating an agent call that sends an email, publishes a record, or initiates a payment can perform the action twice. Such operations require an idempotency key, a result journal, and deduplication at the boundary where the side effect occurs.

AI Feed editorial recommendation: the actions, attempt limits, numerical examples, and fallback choices in this table are practical editorial guidance, not thresholds validated by a controlled study. Teams should verify them against their provider contracts, product requirements, and failure tests.

Set a deadline for the entire operation

A timeout on an individual HTTP call does not limit the user’s total wait. Three attempts of 20 seconds each, with pauses between them, can take considerably longer than a minute even though the user has already left the page. Every request should therefore receive an absolute deadline: a point after which no new attempts begin, a queue no longer waits for a response to a closed connection, and child operations are cancelled.

It is useful to distinguish between:

connection timeout — how long the client waits to establish a connection;
read timeout — how long the client may receive no data;
operation deadline — the maximum duration of the complete user action;
retry budget — the number of attempts and amount of time that may be spent on recovery.

AI Feed editorial example, not a validated study threshold: an interactive assistant could have a total deadline of 12 seconds, one additional attempt only for a temporary error, and at least two seconds reserved for producing a fallback. Background document classification may wait longer, but its job must still have an expiration time: obsolete work should be deleted rather than executed a day later.

Check what the SDK already retries

Some official SDKs automatically retry certain 429 and 5xx responses. The behavior depends on the version and configuration. If an outer service makes three attempts while the SDK makes three attempts inside each one, a single user operation can generate up to nine API calls. This is arithmetic illustrating nested retry multiplication, not a recommended attempt count or a validated threshold.

The OpenAI rate-limits guide recommends limiting both the number of attempts and the total time spent retrying, while accounting for retries built into the SDK. The practical rule is that one component should own the retry policy. Either the application disables automatic SDK retries or includes them in the shared attempt counter and operation deadline.

Retry-After takes precedence over a local formula

If the server provides a valid Retry-After value, the client should not call the service before that interval has elapsed. A small random addition reduces the chance that thousands of clients will wake up at the same moment. If the header is absent or invalid, use bounded exponential backoff with jitter: the pause increases after each failure but has an upper bound.

The algorithm should check before and after sleeping whether the deadline still has time remaining, another attempt is allowed, the circuit breaker is closed, and the user has not cancelled the operation. Infinite retries are unacceptable. Failed requests can themselves consume a per-minute limit, so frequent resubmission can extend an overload event.

Smooth the load before it reaches the provider

Retries can recover from an isolated temporary failure, but they do not regulate the flow of work. Place a bulkhead in front of the AI API to limit the number of calls executing concurrently. This is especially important in fan-out workflows: one file upload can produce hundreds of chunk requests. “Hundreds” is an architecture illustration of possible fan-out, not a universal or validated threshold.

Control not only requests per second but also concurrency, estimated token throughput, and the load attributable to an individual client. Reserve part of the available capacity for interactive actions so that batch indexing cannot displace chat traffic. For confidential or offline workloads, teams can separately evaluate scenarios described in the local AI models guide, but local processing does not eliminate the need for resource management.

A circuit breaker temporarily stops calls to a dependency after a sustained series of failures. It gives the dependency room to recover, but traffic must return cautiously: releasing the entire queue immediately after the breaker closes can create another spike.

Three classes of work require three policies

Class	Primary concern	AI Feed editorial recommendation
Interactive chat or suggestion	Short latency	A short deadline, at most one cautious retry, and a fast fallback
Batch import	Completeness and controlled throughput	A durable queue, checkpoints, bounded fan-out, and retries for individual parts
Background data enrichment	Cost and freshness	Low priority, longer backoff, an expiration time, and deletion of obsolete jobs

AI Feed editorial qualification: these policies and numbers are practical recommendations, not study-validated thresholds. Production values should be calibrated with load tests, cost constraints, user expectations, and provider-specific behavior.

The queue should store a job identifier, deadline, attempt number, priority class, deduplication key, and a secure reference to the input data. A dead-letter queue is useful for investigation, but it should not become an archive of errors that are retried forever. Queues introduce state, delay, and storage requirements; they are a deliberate tradeoff rather than free reliability.

A streaming response cannot be restarted invisibly

With a streaming API, an error can arrive after several paragraphs have already been displayed. An automatic replay can duplicate text, change the output, or invoke a tool again. Both OpenAI and the Gemini API error documentation address errors delivered inside a stream.

The client should retain stream state and distinguish between “nothing has been displayed” and “part of the result has already been consumed.” In the second case, it is safer to mark the response as incomplete and offer the user a continuation. If the model invokes tools, completed calls should be recorded separately and should not run again without a check.

Degradation must be a product decision

A fallback does not necessarily mean switching to another model. Options include a shorter answer without additional analysis, a cached result with its freshness disclosed, placement in a queue, disabling a nonessential AI feature, or an honest error message.

Switching a model or provider changes the format, quality, latency, cost, and safety properties of the result. It should be tested as a separate product branch. The AI API pricing reference can help evaluate the financial tradeoff, but price does not replace compatibility testing.

Observation: providers document temporary restrictions and overload as normal API operating conditions. Forecast qualification: the frequency of these events in a particular product is unknown until load testing and production operation provide evidence. This technical exposure also does not imply that people will “lose their jobs”: a process’s exposure to failure in an AI component describes a system dependency, not an employment forecast. An analysis of AI’s effects on tasks and occupations requires a separate methodology, such as the AI Jobs Impact Map.

What to measure

the share of completed user operations, not merely successful HTTP attempts;
p50, p95, and p99 end-to-end latency, including queueing and retries;
errors by HTTP status, machine-readable code, model, and provider;
attempts per operation and total time spent in backoff;
queue depth and age, including expired and dead-letter jobs;
input and output tokens, plus an estimate of unreported usage when a stream is interrupted;
current concurrency, bulkhead rejections, and circuit-breaker state;
the share of fallbacks, partial streams, and user cancellations.

The percentile labels p50, p95, and p99 are observability categories, not target values or validated thresholds. Pass an internal operation_id, an attempt identifier, and any available provider request ID through the entire chain. Do not record secrets or complete user inputs in logs by default.

Verification methodology

Analysis date: October 19, 2026. This article is based on official OpenAI, Google, and Microsoft documentation available on that date. Its recommendations describe mechanisms rather than fixed quotas.

AI Feed editorial practical checklist — not a validated study protocol or threshold set:

Artificially lower limits in a test environment.
Inject a 429 response with Retry-After.
Inject a 429 response without Retry-After.
Inject a temporary 503 response.
Simulate a response that hangs.
Simulate a network interruption before the first byte.
Interrupt an SSE stream in the middle of delivery.
Confirm that the overall operation deadline is respected.
Confirm that nested retry loops do not occur.
Confirm that the queue cannot grow without a bound or expiration policy.
Confirm that interactive traffic retains reserved capacity.
Confirm that recovery does not release the entire backlog at once.

Run load tests against a production API only with an explicit spending limit, because the test itself can exhaust a quota. The checklist and every practical number above are AI Feed editorial recommendations rather than validated study thresholds. Teams should adapt them to their architecture, provider terms, risk profile, and incident history.

Frequently asked questions

How many times should a request be retried?

There is no universal number. The limit depends on the operation deadline, cost, retry safety, and workload class. An interactive workflow usually needs a very small budget; background jobs may wait longer, but they must also have a finite lifetime. Any example count in this article is an AI Feed editorial recommendation or arithmetic illustration, not a validated study threshold.

Should every 429 response be retried?

No. Determine the machine-readable code and underlying cause. A temporary restriction can permit a delayed retry, while an exhausted daily quota, billing problem, or hard budget limit requires waiting or intervention.

Will a cache solve rate-limit problems?

A cache can reduce average load, but it does not replace throttling. With a cold cache or changing requests, the hit rate can fall sharply, so the system must withstand a large number of cache misses.

When is a queue necessary?

A queue is appropriate when the result does not have to appear in the current HTTP session, the work can be retried safely, and a delay is acceptable. For an interactive response, fast degradation is often preferable to queueing.

What should be implemented first?

AI Feed editorial implementation order: start with an error matrix, an overall deadline, one shared retry budget, and a concurrency limit. Then separate workload classes, add a queue for deferred work, implement observability and a circuit breaker, and test degraded modes. This ordering is practical editorial guidance rather than a study-validated sequence.

Sources and research

AI-assisted article, checked against the primary sources cited above.

Originally published by AI Feed: How to Build a Resilient AI API Integration: Timeouts, Retries, Queues, and Graceful Degradation.

ChatGPT for Teens Gets Study Tools, With College Planner to Follow

AI Feed — Mon, 19 Oct 2026 06:43:10 +0000

OpenAI has added note capture, flashcards and quizzes to ChatGPT for Teens, while a US college-planning tool for grades 10–12 is coming soon.

OpenAI has introduced new study features for ChatGPT for Teens, the default experience for accounts identified as belonging to people under 18. Teens can turn notes into flashcards and interactive quizzes, while a forthcoming College Planner will help US high school students organize applications to four-year colleges.

A continuous capture feature lets users photograph multiple pages of handwritten or printed notes and combine them into one PDF. It is already available on iOS, while Android support remains under development; OpenAI did not specify when the Android version will be released.

Flashcards can be generated from uploaded notes or a chosen topic, saved in a teen’s Library, shuffled and scheduled for later practice. ChatGPT can also convert uploaded notes into interactive quizzes. OpenAI says the system has improved at recognizing requests to create quizzes.

College Planner will be added soon rather than being available at launch with these study tools. Its first version is designed for US students in grades 10–12 who plan to attend four-year colleges. It will consolidate application requirements, deadlines, progress tracking and financial-aid tasks, while answering questions about the process. OpenAI says it is not intended to replace students’ decisions or guidance from counselors, advisers and families.

The initial planner will not cover other countries, two-year institutions, technical colleges or trade schools, although OpenAI says support for those options is planned later. The company is also supporting the expansion of College Advising Corps and the creation of an AI fellowship for advisers.

OpenAI also reported preliminary usage findings. It says teens receiving the experience sent, on average, about 2.7 million more learning-related messages than teens without access. During one week, nearly 1.2 million teens used Learning Visualizations and more than 180,000 used Study Mode. Average daily use was under 15 minutes, and fewer than 2% used ChatGPT for more than three consecutive hours.

Those figures indicate usage, not improved academic performance. The announcement does not state the measurement period for the 2.7 million-message difference, the sample size, how access was allocated or how messages were classified as learning-related. These omissions limit interpretation of the comparison; OpenAI itself calls the findings preliminary and says evaluation will continue with teens, families and educators.

Separately, OpenAI committed to supporting Boston Children’s Hospital’s Digital Wellness Lab and its independent Student Advisory Council for three years. During the 2026–27 school year, approximately 22 students will test tools and advise on teen safety defaults, parental controls, notifications, digital wellbeing and AI literacy. OpenAI also says it will continue testing how protections perform as ChatGPT for Teens evolves.

Sources

openai.com

Originally published by AI Feed: ChatGPT for Teens Gets Study Tools, With College Planner to Follow.

Pollo Agent Uses Three OpenAI Models to Build Ad Campaigns

AI Feed — Sun, 18 Oct 2026 18:03:11 +0000

OpenAI says Pollo Agent turns concepts and references into scripts, storyboards, images and short ads by assigning different models to each task.

Pollo AI’s Pollo Agent uses GPT‑5.6, GPT‑6 Astra and GPT‑Image‑2.5 to turn rough ideas and visual references into storylines, scripts, scene sequences and finished campaign assets, according to an OpenAI profile published on October 8. The product is aimed at creators and marketers producing images, videos and advertising materials.

Each model has a distinct role in the workflow. OpenAI says GPT‑5.6 routes requests, while GPT‑6 Astra handles more demanding narrative development and plans revisions to individual scenes. Pollo AI reports that this allocation reduces the time users spend selecting or switching models by more than 50%, although the source provides no independent test of that figure.

GPT‑Image‑2.5 generates posters, storyboards and other campaign visuals. According to OpenAI, Pollo AI assesses candidate models for quality, visual consistency, stylistic range, speed and cost. The company says its service has more than 26 million users and that 30% of creator sessions begin with a template or workflow based on a trending format.

One featured workflow is Photo to Video Ads within Marketing Studio. A user supplies a product photograph and a prompt, and the tool assembles a short advertisement for social media. In OpenAI’s fragrance example, the system produced a video lasting about 15 seconds, with generated textures, lighting, close-up shots and a staged final composition.

Other showcased concepts include a jewelry campaign placing a model beside a cheetah and a beverage advertisement that turns a bottle photograph into a miniature snowy fantasy landscape. OpenAI attributes these workflows to Pollo Agent and GPT‑Image‑2.5, saying they can reduce production effort, time or cost compared with live shoots. These are publisher-presented examples rather than independently assessed results.

The practical proposition is a single workflow that assigns models without requiring users to choose a separate generator at every stage. However, the source gives no pricing or detailed access terms, and it does not report independent evaluations, latency measurements, accuracy rates, safety results or quantified comparisons of output quality. Claims about savings and effectiveness should therefore be read as company-reported performance, not as an industry benchmark.

Sources

openai.com

Originally published by AI Feed: Pollo Agent Uses Three OpenAI Models to Build Ad Campaigns.

SentinelOne Uses Daybreak to Cut Malware Analysis Time

AI Feed — Sun, 18 Oct 2026 14:32:44 +0000

OpenAI says Daybreak reduced one SentinelOne malware reverse-engineering workflow from four to six weeks to about eight hours, with human oversight.

SentinelOne is using OpenAI Daybreak to accelerate malware reverse engineering, according to OpenAI. In the use case described, work that previously took a top security researcher four to six weeks can be completed in approximately eight hours. This is a result reported by the companies for a specific deployment, not an independent benchmark.

Daybreak gives defenders access to OpenAI models for cybersecurity tasks. SentinelOne characterizes the system’s output in this use case as being at a human level. However, the published account does not specify the evaluation method, describe the malware sample set or quantify how often researchers had to correct the system’s findings.

OpenAI says the time savings allow SentinelOne researchers to conduct deep analysis on more malware samples instead of reserving that work for a small number of especially important threats. The additional capacity also lets them compare findings across a broader collection and search for shared components in malware that initially appears unrelated.

Gregor Stewart, SentinelOne’s chief AI officer, says this wider view can reveal connections between threats and show how those relationships change. Such findings may provide leads for customer investigations and support more proactive defenses. OpenAI also says its cyber capabilities enable a broader group of SentinelOne employees to undertake investigations without relying exclusively on a small pool of specialists.

SentinelOne plans to extend these capabilities into software development so developers can address security issues while writing code. The proposed automation remains bounded by the consequences of a decision: lower-level fixes may be automated, while choices that fundamentally change software architecture should continue to receive human review. OpenAI emphasizes that SentinelOne’s team remains accountable for outcomes.

The broader significance is not merely a faster investigation, but the possibility of examining more samples and testing more links between threats. The reported timeline should not be assumed to apply to every malware family or security team because the account does not break down task complexity or measurement conditions. Human verification therefore remains especially important when findings affect product architecture or customer protection.

Sources

openai.com

Originally published by AI Feed: SentinelOne Uses Daybreak to Cut Malware Analysis Time.

Amp Adds ChatGPT Sign-In and Expands Agents With GPT-6 Astra

AI Feed — Sun, 18 Oct 2026 11:02:42 +0000

Amp has added ChatGPT sign-in, introduced the Puck agent manager and begun testing an Ultrafast mode for longer-running GPT-6 Astra tasks.

Amp announced Sign in with ChatGPT on October 14, giving users of its AI agent more tokens and access to OpenAI models, including GPT-6 Astra. According to OpenAI, the integration is intended to make it easier to delegate larger, longer-running assignments without monitoring every token consumed.

Originally built as an agent for software development, Amp is extending its scope beyond code generation. OpenAI says the service now covers planning, product decisions, writing, design, marketing and post-launch work. OpenAI models account for most tokens used in Amp, while the service selects and recommends models according to users’ needs.

Amp uses GPT-6 Astra for work that requires higher-quality writing and design, OpenAI says, citing documentation, webpages, user onboarding and product discoverability. Amp’s 20-person team also assigns marketing work to agents instead of dedicating an employee to that function. In a separate creative example, CEO Quinn Slack supplied house plans to Astra, which produced a Blender model that his family later adapted into games.

The announcement also introduces Puck, a conversational assistant that launches and manages other agents. OpenAI says Puck can start bug investigations, coordinate several agents and retrieve information from earlier technical conversations. A related feature called Calling Puck uses GPT-Live-1 for real-time voice interaction, including from a watch.

Computing cost remains a material constraint. Some Amp users consume as many as 10 billion tokens per week, according to OpenAI, and concern about additional usage can discourage experimentation. ChatGPT sign-in is meant to ease that constraint through a larger token allowance, but the announcement does not provide the applicable limits, eligibility terms or pricing.

Amp is also testing Ultrafast, a faster mode for GPT-6 Astra. Users currently move between several agent tasks while waiting for results; Amp believes shorter response times could reduce that juggling or leave agents more time to review their own work. OpenAI did not announce general availability, pricing, a release date or performance benchmarks for the mode.

Taken together, the additions position Amp less as a standalone coding assistant and more as a coordinator for extended workflows: users define an objective, while Puck distributes work and recalls earlier context. The announcement illustrates that direction, but without comparative measurements of quality, speed or cost, its practical economics cannot yet be assessed.

Component	Purpose	Status in the announcement
Sign in with ChatGPT	More tokens and access to OpenAI models, including GPT-6 Astra	Integration announced
Puck	Launch and coordinate agents; retrieve earlier discussions	Introduced
Calling Puck	Voice control using GPT-Live-1, including from a watch	Described in the announcement
Ultrafast	Faster operating mode for GPT-6 Astra	Under testing; no release date announced

Sources

openai.com

Originally published by AI Feed: Amp Adds ChatGPT Sign-In and Expands Agents With GPT-6 Astra.

NVIDIA RTX Spark Laptops Go on Sale Starting at $2,399

AI Feed — Sun, 18 Oct 2026 06:17:49 +0000

The first RTX Spark Windows laptops combine Blackwell RTX graphics, a Grace CPU and up to 128GB of unified memory for local AI workloads.

NVIDIA says RTX Spark laptops are now available from $2,399, bringing its new local-AI platform to slim, premium Windows PCs. Initial systems come from ASUS, Dell, HP, Lenovo, Microsoft and MSI, with additional models promised later. The announcement does not identify those future systems or provide a release timetable for them.

The platform centers on a single superchip combining an NVIDIA Blackwell RTX GPU with an NVIDIA Grace CPU. NVIDIA claims up to 1 petaflop of AI compute and up to 128GB of unified memory, positioning the machines for local AI, content creation and gaming. The company also says the laptops include its broader AI platform.

NVIDIA has not published a model-by-model specification list in this announcement. It gives no dimensions, battery figures, regional prices or individual configurations, and it does not specify whether every launch model reaches the stated compute and memory ceilings. The $2,399 figure is therefore a starting point for the product category rather than a universal price across manufacturers and markets.

Up to 128GB of shared memory could let suitably configured systems hold larger local models or workloads than PCs with smaller pools of GPU-accessible memory. However, the 1-petaflop ceiling does not by itself establish token-generation speed, power consumption or performance in a particular application. Meaningful comparisons will require specifications and workload-specific testing for each laptop.

The retail launch included six events worldwide. NVIDIA describes celebrations in Berlin, Beijing, Istanbul, Seoul, San Francisco and Taipei, where attendees saw demonstrations covering gaming, creative workflows and local AI. At Berlin’s XPERION venue, NVIDIA founder and CEO Jensen Huang appeared and gave attendees five RTX Spark laptops bearing his signature.

For developers and AI enthusiasts, NVIDIA points to company-tested recipes, playbooks and other resources for running local AI on RTX Spark. The announcement does not list supported AI models or provide independent benchmark results, leaving buyers to assess the eventual configurations and software performance separately as manufacturers disclose more details.

Category	NVIDIA information
Superchip components	NVIDIA Blackwell RTX GPU and NVIDIA Grace CPU
AI performance	Up to 1 petaflop
Unified memory	Up to 128GB
Initial system manufacturers	ASUS, Dell, HP, Lenovo, Microsoft and MSI
Starting price	From $2,399

Sources

blogs.nvidia.com

Originally published by AI Feed: NVIDIA RTX Spark Laptops Go on Sale Starting at $2,399.

OpenAI and Sullivan & Cromwell Build AI Agreement Analyzer

AI Feed — Sat, 17 Oct 2026 18:02:22 +0000

Agreement Analyzer reviews draft contracts against selected precedents and negotiating positions while leaving every legal decision to lawyers.

OpenAI said on October 15 that Sullivan & Cromwell, working with the AI company, developed Agreement Analyzer to help the law firm’s lawyers identify potential issues in draft agreements. The application combines the firm’s legal materials and established workflows with ChatGPT Enterprise, but does not make final legal decisions.

The project began as a single-user proof of concept built with Codex by Sullivan & Cromwell partner John Savva. OpenAI subsequently worked with the firm to improve output quality, introduce parallel processing and enable several lawyers to use the application simultaneously.

For a review, a lawyer can provide a draft agreement alongside selected precedents, firm and client negotiating positions, and information about market practice. According to OpenAI, Agreement Analyzer examines the material through several review cycles from different perspectives, consolidates the results and flags potential problems. It then produces a memorandum and proposed edits presented as tracked changes.

OpenAI says specialized agents examine deviations from market practice, the allocation of risk, relationships among clauses and ambiguities that might affect litigation. The agents can challenge one another’s conclusions. The resulting findings are separated into those that survive the challenge, those that are dismissed and unresolved matters retained with qualifications.

The company explicitly describes Agreement Analyzer as an aid rather than a substitute for legal expertise. Lawyers can inspect its explanations and supporting precedents, then adapt, reject or accept proposed changes. Its memorandum, findings and tracked edits all remain subject to professional review and judgment.

The approach matters because it aims to expose both the evidence behind a recommendation and the uncertainty surrounding it, rather than presenting every machine-generated finding as equivalent. Categorizing conclusions as sustained, dismissed or unresolved could help lawyers focus their attention on provisions where transaction context and professional judgment are especially important.

The announcement does not provide pricing, a timetable for broader deployment or availability outside Sullivan & Cromwell. OpenAI characterizes wider research and writing assistance as a future opportunity, not a completed product capability. The firm is continuing to explore other tools tailored to specific legal tasks.

Sources

openai.com

Originally published by AI Feed: OpenAI and Sullivan & Cromwell Build AI Agreement Analyzer.

Anthropic Opens Singapore Office and Signs AI Memorandum

AI Feed — Sat, 17 Oct 2026 14:32:49 +0000

The fifth regional office accompanies a Singapore government memorandum covering model evaluation, incident response, research and AI skills.

Anthropic opened its fifth Asia-Pacific office in Singapore on October 14 and signed a memorandum of understanding with the Ministry of Digital Development and Information. The document identifies cooperation on AI safety and security, model evaluation, research and responsible adoption in the public and private sectors.

Ministry Permanent Secretary Kai Fong Chng and Anthropic co-founder Ben Mann signed the memorandum during Singapore International Cyber Week. Anthropic says the parties intend to arrange technical exchanges with the country’s AI safety, security and assurance bodies, addressing evaluation methods, testing standards and responses to incidents. The announcement gives no budget, timetable or mandatory performance measures for this work.

Further proposed activities include sharing research and data from the Anthropic Economic Index, contributing technical and policy expertise as Singapore shapes AI rules, and developing AI capabilities among public officers. Anthropic also plans to support responsible adoption by small and medium-sized businesses. These are prospective fields of cooperation described by the company, not completed programs.

Anthropic reports that, after adjusting for population, Singapore placed second among 121 countries in Claude.ai use, with activity close to six times the level implied by population alone. The company also says more than 1,500 GovTech Singapore engineers and technical employees have taken up Claude Code to update government systems and accelerate software delivery. According to the same announcement, Singapore’s Cyber Security Agency is still examining whether Claude could assist national cyber defense.

The announcement names DBS, GIC, Grab, Sea and Supabase as organizations using Anthropic products. Anthropic says DBS applies Claude to software work and routine productivity under the bank’s security and governance controls, while GIC is introducing it across investment and corporate-services teams. The company adds that regulated organizations can use Claude through Amazon Bedrock with inference conducted in the region.

Anthropic further claims that Grab’s Claude-based BriX sales assistant saves its sales teams approximately 40,000 hours each quarter and identifies new opportunities. This estimate is presented as a vendor-reported customer result; the announcement supplies no independent validation or information about statistical variability. Anthropic has a commercial interest in highlighting adoption of Claude and Claude Code.

A Singapore team may reduce the coordination distance between Anthropic, government bodies and major customers, but the memorandum does not itself announce new AI rules or a specific Claude product change. Its practical importance will depend on whether the broad proposals produce shared evaluation processes, incident-response arrangements and measurable skills or adoption initiatives. Anthropic says the office is open and recruiting.

Area	Proposed work
Safety and evaluation	Exchange technical knowledge on assessing models, test standards and responses to incidents
Research	Examine AI adoption and its effects, with data sharing from the Anthropic Economic Index
Skills and adoption	Build public-officer capabilities and support responsible AI use by small and medium-sized businesses
Regulation	Provide technical and policy expertise as Singapore formulates rules for AI

Sources

www.anthropic.com

Originally published by AI Feed: Anthropic Opens Singapore Office and Signs AI Memorandum.

OpenAI Says ChatGPT Plugin Tripled Stack Overflow for Agents Signups

AI Feed — Sat, 17 Oct 2026 11:04:31 +0000

Stack Overflow used Codex, OpenAI’s models and API, and a ChatGPT plugin to develop and distribute its knowledge platform for AI agents.

OpenAI says registrations for Stack Overflow for Agents—not registrations for the plugin itself—tripled overnight after Stack Overflow introduced a ChatGPT plugin. The product is intended to connect AI agents to verified organizational knowledge, while the plugin enables ChatGPT users to discover and configure the service.

Stack Overflow used Codex, OpenAI’s GPT-6 Astra and GPT-6 Sol models, OpenAI’s API, and the ChatGPT plugin across development and distribution. According to OpenAI, two Stack Overflow engineers gave Codex a development environment and implementation plan, then produced a complete product within a few weeks.

Codex was also used to build connectors for Stack Internal, Stack Overflow’s enterprise product. Alexandra Lato, Stack Overflow’s vice president of product, said a connector would ordinarily require three to six weeks. One was completed with Codex in four days and another in one day. OpenAI describes the first result as a maximum 87% reduction in development time and says Stack Overflow targeted a three-month route to market instead of a year.

For distribution, Stack Overflow created a ChatGPT plugin that OpenAI says involves fewer setup steps than a traditional Model Context Protocol configuration. Lato reported that Stack Overflow for Agents registrations tripled overnight following the plugin’s introduction, while agent creation rose by as much as tenfold. The latter figure is an approximate upper bound based on the latest number she had checked, not a fixed or sustained rate.

OpenAI says the platform routes tasks between GPT-6 Astra and GPT-6 Sol according to reasoning requirements and cost. Stack Overflow currently manages that routing and its related configuration. Lato expressed interest in exploring OpenAI’s Decisions API to automate some of the work, but the source presents this only as a possible future step, not a deployed capability.

Embedding configuration in an environment where users already work with AI agents may reduce setup friction and support adoption. However, the reported development and registration gains come from a single vendor case study and were supplied by the participating companies, so they should not be treated as independent benchmarks. Enterprise connectors still require more human involvement than the initial two-engineer effort, and organization-wide sharing of knowledge learned in individual sessions remains under development.

Stage	Usual duration or level	Reported case result
Product development	Not specified	A few weeks with two engineers using Codex
First enterprise connector	3–6 weeks	4 days; up to an 87% reduction
Next connector	3–6 weeks	1 day
After plugin introduction	Pre-launch baseline	Registrations tripled; agent creation increased by up to 10 times

Sources

openai.com

Originally published by AI Feed: OpenAI Says ChatGPT Plugin Tripled Stack Overflow for Agents Signups.