GPT-6 Astra vs Claude Fable 5.1 vs Gemini 3.1 Pro
Anthropic shipped Claude Fable 5.1 on 1 September 2026 and OpenAI shipped GPT-6 Astra on 3 September. Google's current stable flagship remains Gemini 3.1 Pro; Gemini 3.5 Pro has not shipped and is still in partner testing after missing internal targets, so any comparison naming it is wrong. On published rates, GPT-6 Astra and Claude Fable 5.1 both sit at $10 per million input and $50 per million output, while Gemini 3.1 Pro's preview pricing is $2 and $12 for prompts up to 200,000 tokens: roughly a fifth of the price. Two of the three charge a long-context premium that most coverage omits entirely, with Astra rising to $20 input and $75 output. The sharpest real difference is cache economics. Fable 5.1 cut cache reads by 75% to $0.25 per million, which is 2.5% of its own input price, against 10% for both competitors, and that ratio matters more than headline rates for agents that reuse a large system prompt. Cross-lab benchmark numbers in this post are reported with the harness attached because they are not directly comparable.
Three days that reset the frontier
The first week of September 2026 compressed what is normally a quarter of releases into 72 hours.
On 1 September, Anthropic released Claude Fable 5.1 and Claude Mythos 5.1, documented in a joint system card. On 3 September, OpenAI released GPT-6 Astra. Google's most recent flagship-tier release, Gemini 3.1 Pro, remains its current stable top model, with Flash-tier updates continuing through August.
That last point needs stating plainly because a great deal of published comparison content gets it wrong: Gemini 3.5 Pro has not shipped. Google teased a Pro release alongside the 3.5 Flash launch, but it has not landed. Google DeepMind's product lead has said publicly that 3.5 Pro is in partner testing after the team struggled to meet internal performance goals. The most recent Gemini releases have been Flash-tier: 3.7 Flash on 13 August 2026, and 3.6 Flash and 3.5 Flash-Lite reaching general availability. If a comparison names Gemini 3.5 Pro as a shipped competitor, it was not checked against the changelog.
How we compared
Every price in this article was read directly from the vendor's own pricing page on 6 September 2026 and is linked at the point of use. We have not run these models against a private benchmark suite, and we are not presenting hands-on testing that did not happen. Where a number comes from a lab's own reporting or from third-party coverage, we name the source inline.
One access limitation to disclose: our verification run could not reach Anthropic's pricing page directly. The Fable 5.1 cache figures below come from Anthropic's system card and from VentureBeat's launch coverage, which reported the cache change as the headline commercial fact. Confirm current rates on Anthropic's pricing page before committing a budget.
The pricing table everyone is quoting wrong
Here is the fact that most coverage of these releases has missed entirely: all three models are priced in tiers, and the number being repeated is the cheap one.
| GPT-6 Astra | Claude Fable 5.1 | Gemini 3.1 Pro (preview) | |
|---|---|---|---|
| --- | --- | --- | --- |
| Input, standard tier | $10.00 | $10.00 | $2.00 (≤200k prompt) |
| Output, standard tier | $50.00 | $50.00 | $12.00 (≤200k prompt) |
| Cached input | $1.00 | $0.25 | $0.20 |
| Input, long context | $20.00 | not separately published | $4.00 (>200k prompt) |
| Output, long context | $75.00 | not separately published | $18.00 (>200k prompt) |
| Cache read as % of input | 10% | 2.5% | 10% |
| Context window | 1M+ | 1M | — |
| Max output | 128K | 128K | — |
All figures are per million tokens, read 6 September 2026 from OpenAI's pricing page and Google's Gemini API pricing page. Gemini's rates are listed under preview pricing and Google publishes additional Batch, Flex and Priority tiers, with Batch and Flex at half the standard input and output rates. OpenAI publishes batch pricing at half the standard rate and a Fast mode at 2x.
Two things fall out of this table that are worth more than the headline.
First, the convergence is real but narrower than it looks. OpenAI and Anthropic have landed on exactly the same standard-tier numbers, $10 and $50. That is a genuine and slightly remarkable alignment at the top of the market. But Google is not part of that convergence at all. Gemini 3.1 Pro's standard tier is a fifth of the input price and roughly a quarter of the output price. The frontier is not converging on price; two labs are, and the third is competing on it.
Second, long context is where budgets break. A retrieval pipeline that grows its context payload past a threshold does not fail loudly. It gets more expensive. Astra's long-context tier doubles input to $20 and raises output to $75. Gemini doubles input above 200,000 tokens and raises output by half. If you are building anything that stuffs large documents into a prompt, this row matters more to your bill than any benchmark in this article. Our guide to context engineering covers the techniques for keeping payloads small enough that the question does not arise.
Why cache economics is the real story
The 75% cut to Claude Fable 5.1's cache reads, from $1.00 to $0.25 per million tokens, reads like a modest line item. It is the most consequential pricing change of the week, and the reason is a ratio rather than an absolute.
Cache reads on Fable 5.1 now cost 2.5% of its standard input price. On both GPT-6 Astra and Gemini 3.1 Pro, cache reads cost 10%. For any workload that re-sends a large fixed prefix on every call — a long system prompt, a tool catalogue, a retrieved corpus that does not change between turns — that ratio, not the headline rate, sets your input cost.
Work it through at a 90% cache hit rate, which is unremarkable for a production agent with a stable system prompt:
- GPT-6 Astra: (0.9 x $1.00) + (0.1 x $10.00) = $1.90 effective per million input tokens
- Claude Fable 5.1: (0.9 x $0.25) + (0.1 x $10.00) = $1.23 effective per million input tokens
- Gemini 3.1 Pro: (0.9 x $0.20) + (0.1 x $2.00) = $0.38 effective per million input tokens
Two caveats keep this honest. Cache writes are not free on any of these platforms and are charged above the standard input rate, so a workload with poor cache locality can pay more than the naive rate. And output tokens usually dominate the total bill, so a 35% saving on the input side of an output-heavy workload is a smaller win than the arithmetic above suggests in isolation. Our explainer on when prompt caching actually saves money walks through the cases where it does not.
Where the benchmarks disagree, and why you cannot stack them
Each lab published capability numbers alongside its release. They are useful and they are not comparable, and the second half of that sentence is the part usually dropped.
Claude Fable 5.1 scored 52.6% on Terminal-Bench-Science 0.1, against 24.7% for Fable 5 and 29.0% for Opus 5. That is a same-lab, same-harness comparison across model generations, which makes it one of the more trustworthy numbers of the week: it tells you Anthropic roughly doubled performance on a benchmark it also ran on its predecessors.
GPT-6 Astra posted 72.6% on OSWorld 2.0 for computer use at roughly 47% less time per task than its predecessor, saturated FrontierMath Tier 4 at 97.6% and ExploitBench at 100%, and reached 99.9% on ARC-AGI-3 — the last of these explicitly under OpenAI's provider adapter harness. That qualifier is not a footnote. A harness determines how a model is prompted, how many attempts it gets, and how answers are extracted, and those choices move scores materially.
Artificial Analysis, an independent third party, places GPT-6 Astra's top variant at 55 on its Intelligence Index with output at 71 tokens per second and a cost of $2.57 per Intelligence Index task, alongside a 1M context window and an April 30, 2026 knowledge cutoff. Third-party indices have their own methodological choices, but they at least apply the same ones to every model.
The honest summary: these numbers tell you what each lab chose to optimise and measure. Anthropic is publishing scientific and terminal-agent capability. OpenAI is publishing computer use, frontier mathematics and offensive security. That difference in what gets measured is itself the most informative signal in the set, and it is more useful than any attempt to rank them on a composite. Building your own evaluation harness remains the only way to know how these models perform on your task; our guide to AI evals covers doing that properly.
Access is the new differentiator
The most genuinely new thing about this release week is not capability or price. It is that both frontier labs shipped capability that they will not sell to everyone.
GPT-6 Astra is the first OpenAI model rated Critical under its cybersecurity threshold. Access starts with enterprise customers in OpenAI's Daybreak programme, with paid ChatGPT plans and the API following. That is a staged rollout justified by capability risk rather than by capacity.
Anthropic split the same model in two. Fable 5.1 and Mythos 5.1 are the same underlying model; Fable is the generally available version with production safeguards, and Mythos is offered through restricted-access programmes to vetted cybersecurity and life-sciences organisations that need capabilities those safeguards constrain. Anthropic's system card describes these as the strongest overall cyber capabilities of any model it has released.
Two different mechanisms, one shared conclusion: the frontier is now gated by eligibility, not by price. For a team planning a build, that changes the first question from "can we afford it" to "can we get it," and the answer depends on your organisation rather than your budget. This is a genuine break from the last three years and it is under-discussed relative to its practical impact.
Which should you choose?
If cost is the binding constraint: Gemini 3.1 Pro, by a distance. At $2 input and $12 output it is not competing in the same price band, and for high-volume classification, extraction or summarisation the gap is decisive. Confirm the preview pricing is still current before you build a business case on it.
If you are running a long-lived agent with a large stable prompt: Claude Fable 5.1. The 2.5% cache read ratio is the best in the market and it compounds over a long-running loop, which is exactly the workload it was priced for. The Terminal-Bench-Science result also points at agentic and tool-use competence specifically.
If you need computer use or frontier mathematics: GPT-6 Astra, subject to access. The OSWorld 2.0 and FrontierMath results are the strongest published on those axes. Verify eligibility before designing around it.
If you need predictable long-context costs: check the tier boundaries before anything else. This is the question that will actually surprise you in month two, and it is answered on the pricing page rather than the benchmark table.
If you are choosing an open-weight model instead: that is a legitimate answer for a large share of production workloads, and it is a different comparison. Our open-source LLM comparison covers it, and the ownership change at the main distribution platform, which we cover in our analysis of NVIDIA's acquisition of Hugging Face, does not affect those licences.
Conclusion
The week's headline was two frontier releases in three days. The more useful findings are quieter.
The pricing convergence between OpenAI and Anthropic at $10 and $50 is real, and Google is not participating in it — Gemini 3.1 Pro sits at roughly a fifth the input price and is the value option by a margin that no benchmark difference obviously justifies. Every one of these models charges by context tier, and the widely quoted single price is the cheap tier. Cache economics, not headline rates, decide the bill for agentic workloads, and Anthropic's 2.5% ratio is currently the most aggressive in the market.
And the structural change worth carrying forward: the frontier is now rationed by eligibility. Whether you can use the best available model is becoming a question about who you are rather than what you will pay.
This comparison is an editorial synthesis of vendor pricing pages, published system cards, official changelogs and third-party benchmark indices, all read on 6 September 2026. We did not run private evaluations against these models and make no first-person performance claims. Benchmark figures are reported with their originating harness where the publisher disclosed it, because cross-lab scores are not directly comparable. Model pricing and availability change frequently; verify before relying on any figure here.
Key Takeaways
- Gemini 3.1 Pro is Google's current flagship. Gemini 3.5 Pro has not been released; Google has said it is in partner testing after missing internal performance goals. Comparisons naming 3.5 Pro as a shipped model are incorrect.
- GPT-6 Astra and Claude Fable 5.1 have converged exactly on headline price: $10 per million input tokens and $50 per million output. Gemini 3.1 Pro preview pricing is $2 and $12 for prompts at or below 200,000 tokens.
- Long-context pricing is the omission that matters. GPT-6 Astra charges $20 input and $75 output on long context, and Gemini 3.1 Pro charges $4 and $18 above 200,000 tokens. The single price you see quoted elsewhere is the short-context tier.
- Cache read economics diverge sharply. Fable 5.1's $0.25 per million is 2.5% of its input price after a 75% cut; Astra's $1.00 and Gemini's $0.20 are both 10% of theirs. For agent loops that re-send a large fixed prompt, that ratio drives cost more than the headline rate.
- Astra is the first OpenAI model rated Critical for cybersecurity capability, and access begins with enterprise customers before paid ChatGPT plans and the API. Availability, not capability, may be your binding constraint.
- Anthropic released Fable 5.1 and Mythos 5.1 as the same underlying model with different safeguards: Mythos is restricted-access for vetted cybersecurity and life-sciences organisations. Both labs are now gating frontier capability behind eligibility rather than price.
- Benchmark scores here come from different harnesses and are not interchangeable. Astra's ARC-AGI-3 result is explicitly measured under OpenAI's own provider adapter harness, which is a meaningful qualifier and not a footnote.
Frequently Asked Questions
Which of these three models is cheapest to run?
On published rates, Gemini 3.1 Pro by a wide margin: $2 per million input and $12 output for prompts up to 200,000 tokens, against $10 and $50 for both GPT-6 Astra and Claude Fable 5.1. That pricing is listed as preview, so confirm current rates before committing a budget to it.
Is Gemini 3.5 Pro out yet?
No. As of 6 September 2026 Google's stable flagship is Gemini 3.1 Pro. Google has publicly indicated that 3.5 Pro is being tested with partners after missing internal performance goals, and the most recent Gemini releases have been Flash-tier models including 3.7 Flash on 13 August 2026. Treat any comparison citing a shipped 3.5 Pro as unreliable.
What does the 75% cache read cut actually save?
It matters most for workloads that re-send a large fixed prefix. If 90% of your input tokens are cache hits, Fable 5.1's effective input cost falls to about $1.23 per million versus roughly $1.90 for GPT-6 Astra at the same hit rate. Output tokens are unaffected and usually still dominate the total bill.
Why does GPT-6 Astra cost more for long prompts?
OpenAI publishes two tiers for Astra: short-context input and output at $10 and $50 per million, and long-context at $20 and $75. Google does the same at a 200,000-token boundary. The practical consequence is that a large retrieval payload can double your input cost without any change to your code.
Can I use GPT-6 Astra today?
Not necessarily. OpenAI's rollout starts with enterprise customers in its Daybreak programme, with paid ChatGPT plans and the API following. Astra is the first OpenAI model rated Critical under its cybersecurity threshold, which is the stated reason for the staged release, so check your account's eligibility before designing around it.
What is the difference between Claude Fable 5.1 and Mythos 5.1?
They are the same underlying model. Fable 5.1 is the generally available version with Anthropic's production safeguards applied. Mythos 5.1 is offered through restricted-access programmes to vetted cybersecurity and life-sciences organisations that need capabilities those safeguards normally constrain. You cannot obtain Mythos by paying more.
Are these benchmark scores comparable across the three models?
Not directly. Each lab reports on harnesses it selects and configures, and at least one headline figure, Astra's ARC-AGI-3 score, is explicitly qualified as measured under OpenAI's own provider adapter. Use published benchmarks to identify which capabilities a lab is optimising for, then evaluate on your own task before choosing.
About the Author
Aisha Patel
AI Editorial Desk
AI Editorial Desk · Web3AIBlog
Aisha Patel is a pen name for our AI editorial desk. Posts under this byline are written and reviewed by our team of contributors with backgrounds in machine learning, large language models, AI infrastructure, and applied research. The desk covers frontier model releases, agent architectures, retrieval-augmented generation, on-device inference, and the engineering tradeoffs that matter when shipping AI in production. Every technical claim is verified against primary sources before publication.
