GPT-5.6 series debuts under government oversight, Anthropic warns of AI-driven economic shock, and a new coding benchmark pushes models to their limits.
OpenAI Launches GPT-5.6 Sol, Terra, and Luna in Limited Preview
OpenAI has officially launched its next-generation GPT-5.6 model family, introducing a new naming convention where the number indicates generation and the name identifies durable capability tiers. The family includes Sol (flagship), Terra (balanced), and Luna (fast/affordable). — OpenAI
Sol represents OpenAI's strongest model to date, featuring a new Ultra mode that leverages subagents to accelerate complex work beyond single-agent capabilities. Terra is designed to be competitive with GPT-5.5 while costing 2x less, and Luna offers strong capability at OpenAI's lowest price point. Pricing per 1M tokens runs $5/$30 for Sol, $2.50/$15 for Terra, and $1/$6 for Luna. — The Decoder
The preview also introduces predictive prompt caching with a 30-minute minimum cache life, and a new system card detailing the "most robust safety stack to date" featuring three layers of safeguards. GPT-4.5 has been retired alongside the announcement. A broader rollout is expected "in the coming weeks," with a Cerebras partnership promising up to 750 tokens per second for Sol starting in July.
🔗 OpenAI · The Decoder
GPT-5.6 Rollout Requires US Government Approval "Customer by Customer"
At the request of the U.S. federal government, OpenAI has agreed to initially limit GPT-5.6 access to a small group of trusted partners, with approval granted on a "customer by customer basis." CEO Sam Altman called the arrangement "not our preferred long term model" in an internal memo. — OpenAI · The Information
The move follows the Trump administration's executive order on voluntary AI model review. Despite the "voluntary" framing, Altman received a call from Commerce Secretary Howard Lutnick warning against proceeding without sign-off from more agencies. The situation traces back to Anthropic's forced takedown of its Claude Fable 5 model, creating widespread fear among AI labs of a de facto government licensing regime for frontier AI models. — The Decoder
Altman hopes for broader release "a couple of weeks later," assuming smooth sailing during the preview phase. Talks with the Office of the National Cyber Director and the Office of Science and Technology Policy shaped the phased rollout approach.
🔗 OpenAI · The Information · The Decoder
Anthropic Stops Hiring Junior Engineers, Warns of Economic Shock
Anthropic co-founder Jack Clark revealed that the company no longer hires junior software engineers. "The returns on intuition are much greater than before," Clark said in an interview. Claude handles experimentation at scale — work that used to require large teams of junior researchers. The company now exclusively targets "senior intuition" and experienced hires. — Anthropic · The Decoder
Clark warned of a dangerous economic paradox: AI could produce "far above-trend GDP growth" accompanied by "a spike in unemployment that you typically only see during a recession." He described a scenario where AI multiplies the output of top experts while simultaneously automating entry-level work — a combination no government is prepared for, he said.
The revelation adds to growing concerns about AI's impact on the labor market, especially as coding tools like OpenAI's Codex (now with 4M+ weekly users) and Anthropic's own Claude Code become increasingly capable.
🔗 Anthropic · The Decoder
AI Startup Lindy Ditches Claude for Deepseek, Saving Millions
AI startup Lindy has completely switched from Anthropic's Claude to Deepseek, hosted on US soil by a US company. CEO Flo Crivello told CNBC that AI costs had become "unsustainable," exceeding personnel costs for the 25-person startup. The cost curve "crashed to the ground" after the switch, saving millions. — CNBC · The Decoder
"It's a matter of survival for the business," Crivello said, adding he would switch back if Anthropic cuts prices. The move highlights growing cost pressure on frontier AI labs. OpenAI CEO Sam Altman recently acknowledged that AI cost has become a "huge issue" for companies, especially with agentic systems burning through tokens at unprecedented rates. — Snowflake CTO analysis
A recent analysis by Snowflake's CTO showed that affordable Chinese models like GLM-5.2 are competitive with Opus 4.7 at a fraction of the cost, intensifying the pricing pressure on US AI labs.
🔗 CNBC · The Decoder
MirrorCode Benchmark: AI Models Code Nonstop for 19 Days
Epoch AI and METR have introduced MirrorCode, a new benchmark that requires AI models to recreate complete programs from scratch without access to the original source code. The 25 target programs span Unix utilities, data serialization, bioinformatics, interpreters, cryptography, and compression. — Epoch AI
Claude Opus 4.7 leads the benchmark with a 56% solve rate, successfully reimplementing a 16,000-line bioinformatics toolkit (gotree) in just 14 hours at a cost of $251 — a task that would take a human engineer 2 to 17 weeks. GPT-5.5 follows at 44%, with Gemini 3.1 Pro Preview at 32%. One of the most demanding tasks cost $2,600 to run, with the AI working continuously for 19 days with zero human involvement. — The Decoder
However, none of the tested models can crack the largest tasks. The researchers note that models from a year ago would have scored only about 30%. Epoch AI has open-sourced the scaffold and 22 of the 25 target programs covering 132 task instances across six programming languages.
🔗 Epoch AI (GitHub) · The Decoder
Linux Foundation and 20 Tech Giants Launch Akrites for Open-Source Security
The Linux Foundation has launched Akrites, a coordinated industry initiative to patch security flaws in widely used open-source software before AI-powered attacks can exploit them. Founding members include Amazon Web Services, Anthropic, Cisco, Citi, Google, IBM, JPMorganChase, Microsoft, NVIDIA, OpenAI, Red Hat, the Rust Foundation, Vodafone, and Zscaler. — Linux Foundation · The Decoder
The initiative is driven by a shift in the offensive-defensive balance: modern AI models can scan a large project for vulnerabilities in minutes instead of weeks. Currently, fewer than 5% of validated open-source vulnerabilities have been patched. Akrites establishes a shared Security Incident Response Team (SIRT) as a single point of contact, replacing the current fragmented system where dozens of organizations independently flag the same flaws.
A key feature: when a critical package lacks an active maintainer, Akrites steps in as "maintainer of last resort," shipping the fix itself. Seed funding comes from Alpha-Omega, a directed fund under the Linux Foundation.
🔗 Linux Foundation · The Decoder
OpenAI IPO Likely Delayed to 2027 Over $1 Trillion Valuation Demand
OpenAI is leaning toward delaying its IPO to 2027, according to the New York Times. CEO Sam Altman is pushing for a $1 trillion valuation — up from the company's last private valuation of $730 billion — and has rejected anything below that as a "nonstarter." Advisors recommended the delay due to volatile tech markets and SpaceX's weak post-IPO stock performance. — NYT · The Decoder
The news triggered a 13% drop in SoftBank's stock, as the Japanese mega-investor — one of OpenAI's biggest backers with ~$65 billion invested — had been banking on a quick IPO payoff. OpenAI brought in about $13 billion in revenue in 2025 but continues to post heavy losses. ChatGPT user numbers have stalled at around 900 million, short of the 1 billion target. — Bloomberg
OpenAI has been scaling its B2B business aggressively, with Codex now reaching more than 4 million weekly users — a fivefold increase in three months. The company is betting on enterprise transformation to justify the trillion-dollar valuation.
🔗 NYT · The Decoder · Bloomberg
Top comments (0)