DEV Community: Anders Rasmussen

LLM Pricing Digest: DeepSeek Quietly Drops Two New Flash Models

Anders Rasmussen — Sun, 13 Sep 2026 10:02:23 +0000

LLM Pricing Digest: DeepSeek Quietly Drops Two New Flash Models

Weekly roundup from LLM Price Watch — tracking Claude, GPT, Gemini, DeepSeek, and Grok pricing daily.

This week's data is straightforward: no price changes across any of our five tracked providers. Zero. If you've got a budget locked in for GPT-4o, Claude Sonnet, or Gemini Flash, nothing moved on you.

What did happen is DeepSeek showed up with two new models on OpenRouter, and they're worth knowing about.

What's New from DeepSeek

deepseek/deepseek-v4.1-flash — spotted September 10th.

deepseek/deepseek-v4-flash-vision-exp:batch — spotted September 9th, a day earlier.

Let's talk about what these names actually tell us.

The v4.1-flash model follows DeepSeek's now-familiar naming pattern: a Flash tier is their lower-latency, lower-cost option, comparable in positioning to Gemini Flash or GPT-4o mini. If previous DeepSeek Flash releases are any guide, this is aimed squarely at high-volume, cost-sensitive workloads — think classification, summarization, structured extraction, anything where you're firing off thousands of calls and cost-per-token matters more than raw capability.

The second model is more interesting: deepseek-v4-flash-vision-exp with a :batch suffix. Two things stand out here. First, the vision tag means multimodal input — image understanding on top of text. Second, the exp label means experimental, so treat it accordingly: don't build a production dependency on it yet, but it's worth testing. The :batch variant on OpenRouter typically means asynchronous batch processing at a reduced rate — good for offline pipelines where you don't need a real-time response.

What This Means Practically

If you're currently using DeepSeek for text-only tasks and you've been waiting for a vision-capable option at Flash pricing, this is the moment to run some tests. Vision at Flash cost is a genuinely useful combination — it opens up document parsing, screenshot analysis, and image classification use cases that previously required bumping up to a pricier tier or switching providers entirely.

The batch variant specifically is worth a look if you have any async workloads. Batch endpoints typically offer meaningful discounts over synchronous calls, and if your pipeline can tolerate latency (nightly report generation, bulk document review, etc.), you can often cut costs significantly just by routing to the batch endpoint.

That said, both models are new and one is explicitly experimental. Before committing:

Run your own evals on representative tasks from your actual workload
Check the current pricing on OpenRouter directly — new models sometimes launch at promotional rates that adjust within weeks
Keep an eye on the exp label; experimental models can change, be deprecated, or graduate to stable naming without much notice

The Quiet Week Is Actually Fine

No price changes across Claude, GPT, Gemini, and Grok this week. That's a stable environment for planning. If you've been putting off a cost analysis because prices keep shifting, this is a decent window to benchmark and lock in assumptions.

DeepSeek continues to move fast on model releases. Two models in two days suggests an active release cadence heading into Q4. Whether that means more aggressive pricing pressure on the other providers is something we'll be watching.

I track price changes and new model releases daily across all five major providers. Full live data at llmpricewatch.com.

GPT-6 Astra, Claude Fable, Gemini 3.8: A Busy Week for New LLM Releases

Anders Rasmussen — Tue, 08 Sep 2026 10:37:56 +0000

GPT-6 Astra, Claude Fable, Gemini 3.8: A Busy Week for New LLM Releases

Weekly LLM pricing digest — week of September 5, 2026

No price changes to report this week across the five providers we track (OpenAI, Anthropic, Google, DeepSeek, and xAI). But the model catalog grew noticeably in just a few days, so let's walk through what actually showed up and what it might mean if you're picking a model right now.

OpenAI: GPT-6 Astra lands with a Pro tier and batch variants

The biggest news this week is that OpenAI dropped two new models on September 5th: gpt-6-astra and gpt-6-astra-pro, each with a corresponding :batch variant.

The naming convention here is familiar — the :batch versions are typically lower-cost, asynchronous endpoints suited for workloads where you don't need a real-time response. If you're doing large-scale document processing, evals, or anything that can tolerate a delay, batch is almost always the right call financially.

The split between a base astra and an astra-pro tier suggests OpenAI is continuing the pattern of offering a more capable (and presumably more expensive) variant alongside a leaner everyday option. Worth checking current pricing before assuming which one fits your use case — the gap between base and pro tiers has varied a lot across model generations.

Anthropic: Claude Fable 5.1

Anthropic's new entry, claude-fable-5.1, appeared on September 2nd, again with a batch variant.

The "Fable" name is new in Anthropic's lineup. Whether this sits above, below, or alongside the Sonnet/Opus/Haiku naming structure isn't something our tracker can tell you — we detect models as they appear on OpenRouter, not their internal capability tiers. What we do know is it showed up with batch support from day one, which suggests Anthropic is treating batch as a first-class offering rather than an afterthought.

If you're currently using Claude for any high-volume async workload, it's worth benchmarking fable-5.1 against whatever you're running now.

Google: Gemini 3.8 Flash

gemini-3.8-flash arrived on September 3rd, with batch support included.

The Flash line has consistently been Google's price-performance workhorse — fast, cheap, and good enough for a wide range of tasks. 3.8 continuing that series is a reasonable assumption, though again, actual pricing on our tracker is what should drive your decision, not the name alone.

If you're already using Gemini 2.x Flash for something latency-sensitive or cost-sensitive, this is probably worth a quick test.

xAI: Grok 4.3 Batch

xAI added grok-4.3:batch on September 4th. This appears to be a batch variant of an existing model rather than a completely new release, which fits the pattern of providers progressively rolling out async options for their model lineup.

What this week means practically

Five new model IDs in four days across four different providers, with zero price changes on existing models. A few things stand out:

Batch variants are becoming standard. Every new model this week came with a batch option. If you're not already routing eligible workloads to batch endpoints, you're likely leaving money on the table.
The catalog is expanding faster than pricing is changing. That means the cost landscape for existing models is stable right now, but the choice of which model to use is getting more complex. It's worth revisiting your model selection if you haven't done so recently.
None of these models have a pricing history yet. First-week pricing on a new model sometimes gets revised. It's not unusual to see adjustments in the first few weeks after launch.

We'll be tracking all of these as they settle in. If prices shift on any of the new models — or on anything else across our five providers — it'll show up in next week's digest.

For live pricing on all of these models and more, check LLM Price Watch.

If you want the real story of how this tracker (and its sister site, StackIndex AI) got built — domain traps, design mistakes, the works — it's now a course: StackIndex Academy. First 3 chapters are free, no signup.

GPT-6 Astra, Claude Fable, Gemini 3.8: A Busy Week for New LLM Releases (No Price Changes Though)

Anders Rasmussen — Sun, 06 Sep 2026 10:01:47 +0000

GPT-6 Astra, Claude Fable, Gemini 3.8: A Busy Week for New LLM Releases

Weekly digest from LLM Price Watch — tracking live pricing across OpenAI, Anthropic, Google, DeepSeek, and xAI.

No price changes this week across any of the five providers we track. Zero. But the model catalog? That got noticeably busier. Between September 2nd and 5th, we spotted nine new model IDs across four providers. Here's what showed up and what's worth paying attention to.

OpenAI: GPT-6 Astra (and a Pro tier)

The biggest names to land on our radar are openai/gpt-6-astra and openai/gpt-6-astra-pro, both spotted on September 5th. Both also have batch variants (gpt-6-astra:batch and gpt-6-astra-pro:batch).

The GPT-6 naming is a notable jump — we've been sitting with GPT-4 variants for a long time in practical use. The "Astra" label has been associated with Google's multimodal work in the past, so it's an interesting name choice from OpenAI. The Pro tier suggests a familiar tiered pricing structure is coming, similar to what we've seen with other model families.

Practically speaking: if you're building something today, I'd hold off on committing to these until pricing is confirmed and benchmarks are out. Batch variants being available at launch is a good sign — batch processing has been one of the more reliable ways to cut costs on longer, async workloads.

xAI: Grok 4.3 Batch

x-ai/grok-4.3:batch appeared on September 4th. This is a batch-only variant, no standard inference version detected yet. xAI has been iterating fairly quickly on the Grok 4 line, and adding batch support makes sense if they're targeting workloads where latency isn't critical — think evals, document processing, data pipelines.

Google: Gemini 3.8 Flash

google/gemini-3.8-flash and its batch counterpart showed up on September 3rd. The Flash line from Google has generally been their cost-efficient, faster-response tier — good for high-volume use cases where you don't need the full Gemini Pro/Ultra capability. Gemini 3.8 continuing that Flash tradition makes sense. If past Flash pricing holds, this could be one of the cheaper options in the new generation lineup once pricing is confirmed.

Anthropic: Claude Fable 5.1

anthropic/claude-fable-5.1 and anthropic/claude-fable-5.1:batch appeared on September 2nd, making it the earliest arrival in this week's batch.

The "Fable" name is new for Anthropic — we haven't tracked a Fable line before. It's not clear yet where this sits relative to Sonnet, Haiku, or Opus in terms of capability or price point. The name might suggest a model tuned for narrative or creative tasks, but that's speculation. Worth watching for the official positioning.

What This Week Actually Means

With no price changes, your current cost modeling stays the same. Nothing got cheaper or more expensive in the existing lineup.

The main thing to flag is that four providers dropped new model IDs in a four-day window. That's a lot of surface area appearing at once. A few practical notes:

Don't route production traffic to these yet unless you've tested them. New model IDs appearing in the API doesn't always mean they're fully stable or that pricing is final.
Batch variants at launch is genuinely useful — it signals the providers are thinking about cost-sensitive workloads from day one.
The GPT-6 naming is the most significant signal here. If OpenAI is moving to a 6.x generation label, expect Anthropic and Google to respond with their own positioning in the coming weeks.

I'll be watching for actual pricing data on all of these. Once numbers are confirmed, I'll do a proper comparison.

For live pricing across all tracked models and providers, check llmpricewatch.com.

GPT-6 Astra, Claude Fable, Gemini 3.8: A Busy Week for New LLM Releases

Anders Rasmussen — Sat, 05 Sep 2026 19:04:05 +0000

GPT-6 Astra, Claude Fable, Gemini 3.8: A Busy Week for New LLM Releases

Weekly digest from LLM Price Watch — tracking real pricing data across OpenAI, Anthropic, Google, DeepSeek, and xAI.

No price changes to report this week across our tracked providers. But the model catalog grew significantly — nine new model IDs showed up in our tracker between September 2nd and 5th, 2026, spanning all four of the big players we watch. Here's what appeared and what's worth noting practically.

OpenAI: GPT-6 Astra (and a Pro tier)

The biggest news by name alone: OpenAI dropped two new model families on September 5th — gpt-6-astra and gpt-6-astra-pro, each with a corresponding :batch variant.

The batch variants are the immediate practical signal here. OpenAI has consistently priced batch inference at a meaningful discount (typically around 50% off standard rates), so if you're running non-latency-sensitive workloads — evaluations, document processing, bulk classification — the batch endpoints are worth watching closely once pricing is confirmed.

The split between a base gpt-6-astra and gpt-6-astra-pro suggests a tiered capability structure, similar to how the GPT-4o / GPT-4o-mini split played out. Whether the Pro variant is meaningfully better for reasoning-heavy tasks or just a marketing distinction remains to be seen. I'd hold off on routing production traffic until pricing and benchmark data are clearer.

xAI: Grok 4.3 Batch

On September 4th, grok-4.3:batch appeared — notably, just the batch variant, with no corresponding standard model showing up in our tracker this week. That's an interesting release pattern. It could mean the base grok-4.3 was already present in the catalog, or that xAI is leading with batch access before a broader rollout. Either way, if you're already using Grok models for async workloads, this is worth checking against your current setup.

Google: Gemini 3.8 Flash

September 3rd brought gemini-3.8-flash and its batch counterpart. The Flash line from Google has generally been their price-performance sweet spot — faster and cheaper than Pro variants, with quality that's often sufficient for structured output tasks, summarization, and classification.

If the 3.8 Flash follows the same pricing trajectory as earlier Flash models, it could become a solid default for high-volume, cost-sensitive applications. The batch variant showing up simultaneously with the standard model is a good sign — it suggests Google is treating batch access as a first-class feature rather than an afterthought.

Anthropic: Claude Fable 5.1

The most intriguing naming this week: claude-fable-5.1 appeared on September 2nd, again with a batch variant. "Fable" isn't a naming convention we've seen from Anthropic before — their previous lines have been Haiku, Sonnet, and Opus. A new name likely means a new positioning, possibly a specialized or fine-tuned variant rather than a general-purpose tier.

Without more documentation from Anthropic, it's hard to know where Fable sits in their lineup — whether it's meant to replace something, complement the existing tiers, or target a specific use case. The 5.1 version number suggests it's not a ground-up new model but an iteration of something in the Claude 5 family.

What This Means Practically

A week with nine new model IDs and zero price changes is a week for watching, not necessarily acting. A few things I'd suggest:

Don't rush to migrate to any of these until pricing is confirmed and published. New model IDs on OpenRouter don't always come with finalized pricing on day one.
Batch variants are worth bookmarking across all four providers. If your workload tolerates async processing, batch endpoints typically offer the best cost efficiency available.
The Anthropic "Fable" naming is unusual enough to keep an eye on — it may signal a meaningful capability shift or specialization that affects how you'd evaluate it against Sonnet or Haiku.

I'll be tracking pricing as it settles across all nine of these models in next week's digest.

For live pricing and model tracking across Claude, GPT, Gemini, DeepSeek, and Grok, visit llmpricewatch.com.

New Batch Models From DeepSeek, Google, and OpenAI: LLM Pricing Digest

Anders Rasmussen — Sun, 30 Aug 2026 10:01:38 +0000

LLM Price Watch Weekly Digest

No price changes this week across the five providers we track — Claude, GPT, Gemini, DeepSeek, and Grok all held steady. But there was a quiet flurry of new model IDs showing up on OpenRouter, all of them batch variants, and worth knowing about if you're running async workloads.

Here's what we spotted on August 29th.

DeepSeek: Two New Batch Models

DeepSeek added two new batch-mode models:

deepseek/deepseek-v4-pro-0813:batch
deepseek/deepseek-v4-flash-0731:batch

The naming convention tells you a few things. The -pro vs -flash split mirrors what we've seen before — pro for heavier reasoning tasks, flash for faster and cheaper throughput. The date suffixes (0813, 0731) suggest these are specific checkpoints, which is useful if you care about reproducibility or are comparing outputs across versions.

Batch mode generally means you submit a job and get results back asynchronously, usually at a lower cost per token than the real-time endpoint. If you're doing large-scale data processing, eval runs, or document analysis where latency doesn't matter, batch is almost always the right call. Worth checking whether DeepSeek's batch pricing lands meaningfully below their synchronous rates.

Google: Gemma 4 31B Batch

google/gemma-4-31b-it:batch

Gemma 4 at 31B is a reasonable mid-size open-weights model. The it tag means instruction-tuned. Seeing it show up as a batch endpoint on OpenRouter suggests Google is continuing to expand the Gemma family's API availability beyond just the hosted Gemini models. For teams that want Google's open-weights lineage without committing to Gemini pricing, this is worth a look — especially for batch jobs where you can afford to wait.

OpenAI: Two "OSS" Models

These two caught my eye:

openai/gpt-oss-120b:batch
openai/gpt-oss-20b:batch

The oss label is interesting. OpenAI hasn't broadly publicized models under that naming scheme, so I'm not going to speculate too much about what's behind it. What's observable is that there's a 120B and a 20B variant, both batch-only for now, both appearing on OpenRouter. The size gap between them is large — 120B puts it in flagship territory, 20B is more of an efficient workhorse. If these are genuinely open or open-weight models from OpenAI, that would be notable, but I'd wait for official documentation before drawing conclusions.

For now, treat them as "spotted in the wild" and monitor whether pricing and model cards show up.

What This Week Means Practically

If you're actively choosing models right now:

Batch is underused. Most developers I talk to default to synchronous endpoints even when their use case has no real latency requirement. If you're running nightly pipelines, processing large document sets, or doing offline evals, switching to batch can cut costs significantly. This week's additions give you more batch options across providers.

DeepSeek continues to be prolific. Two new models in one week, both with dated checkpoints, suggests active development. If you're relying on DeepSeek for production workloads, pinning to a specific checkpoint (like 0813) is probably smarter than using a floating alias.

OpenAI's OSS naming is worth watching. No firm conclusions yet, but if these turn out to be open-weight models with competitive pricing, that changes the calculus for teams currently choosing between self-hosting and API access.

No pricing drama this week — just new surface area to explore. I'll flag it here if anything changes on the cost side.

We track live pricing and new model appearances for Claude, GPT, Gemini, DeepSeek, and Grok at llmpricewatch.com.

LLM Pricing This Week: DeepSeek Quietly Drops a Vision Model

Anders Rasmussen — Sun, 23 Aug 2026 10:02:11 +0000

LLM Pricing This Week: DeepSeek Quietly Drops a Vision Model

Weekly digest from LLM Price Watch — we track live pricing for Claude, GPT, Gemini, DeepSeek, and Grok so you don't have to.

It was a quiet week on the pricing front. No cuts, no hikes, nothing moved across the five providers we track. If you're mid-project and budgeting around current rates, your spreadsheet is still valid.

The only thing worth noting this week is a new model showing up in the wild.

New: deepseek/deepseek-v4-flash-vision-exp

On August 22nd, our tracker spotted deepseek/deepseek-v4-flash-vision-exp appearing on OpenRouter for the first time. It's sitting under DeepSeek's lineup, and the name tells you a few things:

v4 — this is positioned as their fourth-generation base
flash — expect a speed/cost optimized variant rather than a full-capability flagship
vision — multimodal, so it can handle image inputs alongside text
exp — experimental, meaning DeepSeek hasn't called this production-ready yet

The exp tag is worth paying attention to. Experimental models from any provider can change behavior, pricing, or disappear entirely without much notice. If you're evaluating this for something that needs to stay stable, treat it as a preview rather than a dependency.

That said, DeepSeek's "flash" tier has generally meant aggressive pricing in the past, and a vision-capable model in that tier is interesting. A lot of use cases — document parsing, receipt extraction, basic image captioning — don't need the heaviest multimodal model available. If DeepSeek prices this the way they've priced their other flash variants, it could be worth benchmarking against GPT-4o mini or Gemini Flash for vision tasks.

We don't have confirmed pricing locked in for this model yet since it just appeared. We'll update the tracker as that settles.

The Bigger Picture This Week

Honestly, a week with no price changes and one experimental model drop is pretty normal. The pace of model releases has been fast enough over the past year that it's easy to assume something significant happens every week, but sometimes the answer is just: nothing moved, here's what's new.

A few things I'd keep in mind heading into next week:

DeepSeek continues to ship fast. Whether or not deepseek-v4-flash-vision-exp turns into a production model worth using, the cadence from DeepSeek has been consistent. They tend to iterate in public, so experimental models often do graduate to stable relatively quickly.

Vision is getting commoditized. A year ago, image input was a premium feature. Now it's showing up in flash/lite tiers across providers. If you're paying for a heavier model primarily because you need vision and assumed cheaper options didn't have it, it's worth re-checking the landscape.

No price changes this week doesn't mean no price pressure. Several providers have cut prices in 2025, and competition hasn't slowed. Stable prices one week doesn't signal a plateau — it just means nothing happened this particular week.

That's the week. One new experimental model from DeepSeek, everything else held steady. If you want to keep an eye on pricing as it shifts — especially for DeepSeek's new addition once pricing confirms — the tracker is running daily checks.

→ llmpricewatch.com

New Models From Google, DeepSeek, and xAI This Week — LLM Price Watch Digest

Anders Rasmussen — Sun, 16 Aug 2026 10:01:33 +0000

LLM Price Watch Weekly Digest — Week of August 14, 2026

No price changes to report this week across Claude, GPT, Gemini, DeepSeek, and Grok. That's actually notable in itself — it's been a period of relative pricing stability. But it wasn't a quiet week model-wise. Four new models showed up in our tracker across three providers, and they're worth knowing about if you're evaluating what to build on.

Google: Gemini 3.7 Flash (and a Batch Variant)

The two biggest additions are google/gemini-3.7-flash and google/gemini-3.7-flash:batch, both first spotted on August 14th.

Gemini Flash has been Google's go-to for cost-efficient, lower-latency tasks — the kind of workloads where you need volume without burning through budget. The 3.7 iteration landing now suggests Google is continuing to iterate on that line in parallel with their heavier models.

The batch variant is the one I'd pay attention to if you're running anything offline — document processing, evals, bulk classification, that sort of thing. Batch endpoints typically come with a meaningful price discount in exchange for higher latency, so if your use case doesn't need a real-time response, it's worth checking the pricing before defaulting to the standard endpoint. We'll have the confirmed pricing posted on LLM Price Watch as soon as it's stable in the tracker.

DeepSeek: v4 Pro 0813

deepseek/deepseek-v4-pro-0813 showed up on August 13th. The 0813 suffix is a date stamp — a common convention DeepSeek uses to version checkpoint releases. This appears to be a refreshed checkpoint of DeepSeek v4 Pro rather than an entirely new architecture.

DeepSeek has been one of the more interesting providers to watch from a price-to-performance standpoint over the past year. Their models have consistently undercut comparable Western models on cost. If you've been using an earlier v4 Pro checkpoint, it's worth running a quick eval against this one to see if quality has shifted. Sometimes these point releases are minor; sometimes they're not.

xAI: Grok 4.6

x-ai/grok-4.6 also appeared on August 13th. xAI has been incrementing Grok's version numbers at a reasonable clip. Grok 4.6 sits between major releases, so this reads like a refinement update — likely improvements to instruction following or reasoning rather than a fundamentally different model.

Grok has carved out a niche for users who want strong general-purpose performance and are already in the xAI ecosystem. Pricing on this one we'll confirm as the data firms up.

What This Week Means Practically

If you're in the middle of a model selection decision right now, here's how I'd think about these additions:

High-volume, latency-tolerant workloads: Look at gemini-3.7-flash:batch once pricing is confirmed. Batch endpoints are consistently underused by developers who could benefit from them.
Cost-sensitive inference: Keep an eye on the DeepSeek v4 Pro 0813 pricing. DeepSeek tends to be aggressive here and a checkpoint update sometimes comes with a price adjustment.
General-purpose API work: Grok 4.6 is worth a quick benchmark if you're already evaluating xAI models.

The lack of price changes this week means your existing cost projections are still valid — no surprises there. But new model versions mean your performance baselines might shift if providers quietly improve quality at the same price point.

I track all of this daily at LLM Price Watch — live pricing for Claude, GPT, Gemini, DeepSeek, and Grok in one place, with alerts when something changes. Worth bookmarking if you're making cost-sensitive model decisions.

Two New Models Just Dropped: DeepSeek V4 Pro and Grok 4.6 — LLM Pricing Digest

Anders Rasmussen — Thu, 13 Aug 2026 14:47:08 +0000

Two New Models Just Dropped: DeepSeek V4 Pro and Grok 4.6 — LLM Pricing Digest

Weekly roundup from LLM Price Watch — we track live pricing across Claude, GPT, Gemini, DeepSeek, and Grok so you don't have to.

No price cuts or hikes to report this week across our tracked providers. What we did catch: two new model IDs appearing in the wild on August 13th — one from DeepSeek and one from xAI. Here's what we know so far.

DeepSeek V4 Pro (0813)

Model ID: deepseek/deepseek-v4-pro-0813

DeepSeek continues their habit of date-stamping model releases, which is actually a useful practice — it makes it easy to tell at a glance which checkpoint you're running. The "V4 Pro" naming suggests this sits above their previous V3 line, though we don't yet have benchmark comparisons or confirmed context window specs to share. The 0813 suffix matches the date it first appeared in our tracker.

DeepSeek has consistently been one of the more cost-competitive options in our index, so if V4 Pro follows that pattern, it's worth watching. We'll update pricing on the site as soon as it's confirmed.

Practical take: If you're currently running DeepSeek V3 in production and cost-efficiency is a priority, keep an eye on how V4 Pro prices in. DeepSeek's previous generational jumps have held the line on price while improving capability — but don't assume that until the numbers are public.

Grok 4.6

Model ID: x-ai/grok-4.6

xAI is pushing another point release with Grok 4.6, spotted the same day as the DeepSeek drop. Grok 4 was already in our tracker, so this is an incremental update rather than a major new generation. Point releases from xAI have sometimes come with quiet capability improvements or adjusted rate limits, so it's worth checking if you're an active Grok user.

Pricing for 4.6 isn't confirmed in our system yet. Grok 4 has sat at a premium compared to some of the DeepSeek options, so if 4.6 comes in at the same price point, the question is whether the incremental improvements justify staying on that tier versus alternatives.

Practical take: If you're already using Grok 4 via API, test 4.6 on your actual use cases before migrating. Point releases don't always move the needle on the tasks that matter most to your specific workload.

The Bigger Picture This Week

Two new models, zero price changes. That's actually a notable signal in itself. We're in a period where the model release cadence remains fast, but the pricing floor seems to have stabilized — at least for this week. The race to the bottom on token pricing that defined much of early 2025 appears to have leveled off, at least temporarily.

For anyone building on top of these APIs right now, the practical advice is the same as always: don't hard-code model names if you can avoid it, because the deepseek-v4-pro-0813 you integrate today may be superseded by a -0901 variant before your next sprint is done. Abstract your model selection layer where possible.

We'll be watching both of these new models closely over the coming days as pricing details get confirmed and early benchmarks surface.

I track these changes daily at *llmpricewatch.com** — live pricing for Claude, GPT, Gemini, DeepSeek, and Grok, updated automatically. If you want alerts when something changes, the site has you covered.*

I built a pricing API for LLMs — then realized the real users might not be human

Anders Rasmussen — Sat, 08 Aug 2026 06:02:20 +0000

Klikk her: LLM Price Watch started as a simple problem: comparing per-token pricing across Claude, GPT, Gemini, DeepSeek, and Grok meant opening five pricing pages and doing the math by hand every time a new model dropped. So I built a calculator. Then I built an API behind it. Then I noticed something about who was actually going to call that API.

The obvious version

The first version of the API was exactly what you'd expect:

GET /v1/models — every tracked model with current pricing
GET /v1/models/:id — a single model
GET /v1/calculate?model=X&input_tokens=N&output_tokens=N — cost for a specific call

Straightforward. A human developer hits /calculate, gets a number, builds their cost estimate into a dashboard somewhere. Done.

The part that changed the design

The actual differentiator turned out to be a fourth endpoint: GET /v1/recommend?use_case=X. Instead of just returning prices, it returns a recommendation — which model fits a given use case (long-document summarization, high-volume classification, coding assistance, customer support) based on both price and the editorial analysis already written for the comparison pages on the site.

Once that endpoint existed, the actual audience for this API stopped being "a developer building a cost dashboard" and started including something else: AI agents doing their own tool selection at runtime. An agent framework deciding which model to route a task to doesn't want to read a blog post — it wants a structured answer to "given this use case, what should I use, and what will it cost me." That's a tool call, not a page view.

That reframing changed a few concrete decisions:

CORS is wide open on purpose. This isn't an API with a dashboard in front of it — it's meant to be called directly from wherever the calling code lives, including client-side agent code.
No API key required (for now). Every bit of friction between "an agent wants this data" and "an agent gets this data" is friction against the actual use case. A paid tier with rate limits is the natural future step once real usage justifies it, but gating from day one would have defeated the point.
/recommend returns reasoning, not just a model name. An agent — or the person who built it — needs to know why, not just what, or the recommendation is a black box nobody trusts enough to actually wire into a decision.

The unglamorous half of this

None of that matters if the numbers are wrong. Pricing data for five providers was verified directly against each provider's own official pricing page, not pulled from a third-party aggregator — aggregators lag, and stale pricing data is worse than no data for something meant to inform actual spend decisions. Updates are still a manual snapshot for now; an auto-refreshing worker is the obvious next step once there's enough usage to justify the engineering time.

The part I didn't expect

The API ended up feeding the other site I run, StackIndex AI — its cost calculator page calls this API live, gets a model recommendation plus a cost estimate, and returns it inline. Two separate sites, same backend, CORS'd across domains, tested end-to-end. I didn't plan the API to be reusable infrastructure when I built it — it just turned out that "structured, agent-callable, reasoning-included" is a useful shape for more than one problem.

If you're building something similar: design the response for a reader who can't ask a follow-up question. That constraint does more for API design than almost anything else.