GPT-6 Astra vs Claude Fable 5.1: Benchmarks, Price and Early Feedback

What launched, and when

OpenAI released GPT-6 Astra on September 3, 2026, while Anthropic announced Claude Fable 5.1 on September 1. Both are real, current frontier models; neither name is a placeholder or a substitute for another product.

OpenAI says Astra is rolling out first to a limited set of organizations and then to ChatGPT Plus, Pro, Business and Enterprise users, with API availability through OpenAI plus Microsoft Azure and AWS Bedrock. Its standard API list price is $10 per million input tokens and $50 per million output tokens.

Anthropic says Fable 5.1 is available to Pro, Max, Team and Enterprise users and through the Claude Platform and major cloud marketplaces. Its base API list price is also $10 per million input tokens and $50 per million output tokens. Anthropic additionally lists cache reads at $0.25 per million tokens and says the cache-price reduction can materially lower costs for agentic workloads.

Sources: OpenAI GPT-6 Astra and Anthropic Claude Fable.

The cleanest independent comparison currently favors Fable on broad intelligence

Artificial Analysis evaluated both releases close to launch. Its max-effort Intelligence Index result is 66 for Claude Fable 5.1 and 61 for GPT-6 Astra. The same evaluator reports a 1-million-token context window for both model families.

That comparison needs an important qualification: Artificial Analysis says its Fable 5.1 run used Anthropic's default server-side fallback, with about 4% of output tokens routed through fallback models on the Intelligence Index. It therefore measures the production-style Fable service configuration rather than an isolated underlying model in every safety-sensitive case.

Artificial Analysis also describes a different picture for coding-agent work: Astra makes a large gain over GPT-5.6 Sol and is competitive with the leading Claude coding configurations on its Coding Agent Index. That is more useful than treating one aggregate intelligence score as a universal ranking.

Sources: Artificial Analysis on Fable 5.1 and Artificial Analysis on GPT-6 Astra.

Vendor benchmark claims: useful, but not independent

OpenAI's Astra release reports several direct comparisons using its stated evaluation setups. In the coding table, OpenAI reports 57.9% for Astra versus 55.8% for Fable 5.1 on Terminal-Bench 4.0; 74.1% versus 67.4% on DeepSWE v1.1; and 64.5 versus 63.6 on FrontierCode 1.1 Extended. On Terminal-Bench Science 0.1, OpenAI reports 64.6% for Astra and 52.6% for Fable 5.1.

The same OpenAI page shows the reverse on some broad reasoning measures: its reproduced Artificial Analysis Intelligence Index row is 61.2 for Astra versus 65.7 for Fable 5.1, and Humanity's Last Exam with tools is 57.2% for Astra versus 65.0% for Fable 5.1.

These are vendor-published results. OpenAI states that scores are maxima at an effort setting and that its research/API evaluation environment can differ from production ChatGPT. Anthropic also warns that production safeguards affect Fable 5.1 benchmark outcomes and that some safety-sensitive requests may route to other Claude models. Those differences matter when interpreting small score gaps.

SWE-bench Verified is not SWE-bench Pro

SWE-bench Verified and SWE-bench Pro should not be merged into one coding score.

SWE-bench Verified is a human-validated subset of 500 SWE-bench instances built from real GitHub issues. The official SWE-bench site describes it as a 500-task subset and provides a standardized mini-SWE-agent "bash-only" setting for more comparable model runs. However, OpenAI stopped reporting SWE-bench Verified in February 2026 after finding evidence of benchmark contamination and exposure to benchmark problems or solutions during training. OpenAI's conclusion was that the benchmark no longer cleanly measures frontier coding progress.

SWE-bench Pro is a separate benchmark developed by Scale. Scale describes 1,865 total tasks across 41 repositories: a 731-task public set, a 276-task private set and an 858-task held-out set. It was designed for longer-horizon, more realistic repository work and uses reproducible environments plus human verification.

But SWE-bench Pro also has a major caveat. In July 2026, OpenAI published an audit estimating that roughly 30% of SWE-bench Pro tasks have material issues. That means a Pro score should be treated as one piece of evidence, not as a definitive coding-agent ranking.

As of this review, Scale's standardized public leaderboard does not provide a directly comparable row for both GPT-6 Astra and Claude Fable 5.1. Third-party boards that list Fable 5.1 around the low-80s on "SWE-bench Pro" can be using different scaffolds or provider-reported settings. We therefore do not rank Astra against Fable 5.1 using those rows.

Sources: SWE-bench Verified, Scale SWE-bench Pro public leaderboard, OpenAI on SWE-bench Verified contamination, and OpenAI's SWE-bench Pro audit.

Agents, computer use and practical tradeoffs

Astra's strongest launch story is action inside software. OpenAI reports 72.6% on its OSWorld 2.0 offline setup and says the model completed those simulated tasks in about 47% less time per task than GPT-5.6 Sol. OpenAI also reports 41.4% on AutomationBench, compared with 31.4% for Fable 5.1 in its table.

Fable 5.1's strongest product story is long-running knowledge and coding work. Anthropic positions it for multi-hour and multi-application agents, large codebase changes, code review and asynchronous projects. Its launch page says the model can run unattended as a managed agent and can use vision to check its own coding work.

The cost decision is not captured by the identical $10/$50 headline token prices. Cache policy, reasoning effort, tool calls, retry behavior, output length and provider routing can dominate the bill for long agent sessions. Artificial Analysis measured Fable 5.1 as more verbose and more expensive per Intelligence Index task at max effort than Astra, while Anthropic's much cheaper cache-read rate can make repeated-context workflows substantially cheaper than its headline rate suggests.

Early public feedback is mixed and highly biased

Public feedback is still too young to call a consensus. A Reddit user who analyzed 22,022 of their own Fable 5 and Fable 5.1 API calls reported more tokens per prompt with 5.1 but a lower estimated dollar cost per prompt after applying the new cache-read pricing. Other Fable threads praise coding capability while complaining that subscription usage appears to deplete quickly. These are user anecdotes, not controlled evaluations, and the people posting are a self-selected sample.

Astra discussion shows the same early-launch problem. One highly active developer thread contained both excitement about apparent coding capability and skeptical replies from people who had not yet received the staged rollout. Another first-use report complained about rapid usage consumption on a small task. Again, those reports do not establish average latency, reliability or cost.

Search-engine-accessible X posts were not reliable enough in this review to quote or treat as measured evidence. Rather than inventing a social consensus, we are leaving X sentiment unscored until reproducible or directly attributable posts can be verified.

Which one looks better right now?

For broad benchmarked intelligence under the independent Artificial Analysis setup, Claude Fable 5.1 currently has the stronger result. For computer use and several coding/automation evaluations published by OpenAI, GPT-6 Astra has a strong launch case, including narrow wins over Fable 5.1 in OpenAI's Terminal-Bench 4.0, DeepSWE and AutomationBench comparisons.

That is not the same as declaring an overall winner. The two vendors use different harnesses and safeguards, independent benchmark coverage is incomplete, SWE-bench Verified has contamination concerns, SWE-bench Pro has task-quality concerns, and access is still rolling out.

A practical evaluation should test the exact workflow: repository size, tool permissions, context reuse, latency tolerance, retry policy, cache hit rate and required safety behavior. For a production coding agent, cost per successfully completed task and regression rate are more actionable than a single headline benchmark percentage.

Confidence and what to watch next

Confidence is high on model identity, launch dates, list pricing and stated availability because those come from vendor pages. Confidence is medium on cross-model benchmark comparisons because harnesses, safeguards and effort settings differ. Confidence is low on community-wide sentiment because launch-week feedback is sparse, self-selected and affected by staged access.

The next useful evidence will be standardized third-party runs of both models under the same current coding-agent harness, especially on contamination-resistant or freshly constructed tasks, plus measured latency and cost-per-success over repeated real repositories.