Key Takeaways
- Neither model wins outright. Claude Fable 5.1 leads on sustained agentic coding, knowledge work and business automation; GPT-6 Astra leads on scientific terminal tasks, abstract reasoning and vision-to-code.
- Both cost exactly the same per token: $10 per million input and $50 per million output.
- Anthropic cut Fable 5.1 cache reads by 75% to $0.25 per million, lowering typical workload costs by around 25% and heavily agentic workloads by up to about 45%.
- Fable 5.1 shipped 1 September 2026; GPT-6 Astra followed on 3 September in limited preview, with general paid access the next day.
- The two companies report conflicting figures for each other’s models, and several headline scores come from evaluation harnesses rather than the model alone.
- For long unattended runs, partner testing favours Fable 5.1; for desktop and browser breadth at speed, Astra’s reworked Codex harness posts the larger gains.
Fable 5.1 vs GPT-6 Astra – artistic impression. Image source: Alius Noreika / AI
The Short Answer on Task Execution
If your work involves long-running agentic coding, multi-step business workflows or knowledge tasks that must survive hours without supervision, Claude Fable 5.1 currently executes better. It posts 55.8% on Terminal-Bench 4.0 against 52.3% for Claude Opus 5 and 37.3% for GPT-5.6 Sol, leads AutomationBench at 31.4%, and tops CursorBench 3.2.0 at 73.4%. Partner testing supports the benchmark picture: Browserbase reported Fable 5.1 completing 82% of tasks on its hardest browser-agent benchmark in roughly ten minutes each, against 74% for Opus 5 and 57% for Fable 5, while burning fewer tokens than either.
If your work is scientific research at the command line, mathematical reasoning, or converting visual input into working code, GPT-6 Astra executes better. OpenAI reports 64.6% on Terminal-Bench Science against Anthropic’s 52.6% for Fable 5.1, a 98.6% score on ARC-AGI-3, 97.6% on FrontierMath Tier 4, and 95.9% on the BenchCAD Vision2Code subset against 84.3% for Fable 5.1. The pricing is identical at $10 and $50 per million tokens, so the decision rests on task shape rather than budget.
Launch Timing and Availability
Anthropic released Claude Fable 5.1 on 1 September 2026 alongside Claude Mythos 5.1, the same underlying weights behind more permissive safeguards for vetted cyber defenders and life scientists. Fable 5.1 is generally available as claude-fable-5-1 on the Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud and Microsoft Foundry, plus the Claude apps on Pro, Max, Team and Enterprise. Our guide to the Claude Mythos model family covers the two-tier structure in detail.
OpenAI announced GPT-6 Astra two days later, on 3 September, initially to enterprise customers holding Daybreak access. General paid availability followed on 4 September across ChatGPT Plus, Pro, Business and Enterprise, the OpenAI API and AWS. The API model ID is gpt-6-astra, with a 1,050,000-token context window, 128,000 maximum output tokens and an April 2026 knowledge cutoff.
Head-to-Head Benchmark Table
The table below separates what each vendor published. Cross-vendor figures carry caveats, noted underneath.
| Benchmark | Claude Fable 5.1 | GPT-6 Astra | Reported by |
|---|---|---|---|
| Terminal-Bench Science 0.1 | 52.6% | 64.6% | Anthropic / OpenAI separately |
| Terminal-Bench 4.0 | 55.8% (60.9% as Mythos 5.1) | Not published | Anthropic |
| DeepSWE v1.1 | 67.4% (OpenAI’s figure) | 74.1% | OpenAI |
| CursorBench 3.2.0 | 73.4% | Not published | Anthropic |
| OSWorld (desktop control) | 77.9% partial / 41.7% strict | 72.6% on V2-Offline | Different releases, not comparable |
| GDPval-AA v2 (knowledge work) | 1853 Elo | Not published | Anthropic |
| AutomationBench | 31.4% | Not published | Anthropic |
| Humanity’s Last Exam | 60.9% no tools / 65.0% with tools | Not published | Anthropic |
| ARC-AGI-3 | Not published | 98.6% | OpenAI |
| FrontierMath Tier 4 | Not published | 97.6% | OpenAI |
| BenchCAD Vision2Code | 84.3% | 95.9% | OpenAI |
Why the Numbers Disagree
Three caveats matter more than any individual score, and readers comparing these results should weigh them carefully.
First, the DeepSWE gap is smaller than OpenAI’s chart suggests. The public DeepSWE leaderboard places Gemini 3.8 Flash and Claude Opus 5 at 74% and GPT-5.6 Sol at 73%, with overlapping uncertainty ranges that establish no clear leader. OpenAI’s own chart used a 67.4% figure for Fable 5.1 and excluded Meta’s Muse Spark 1.3, which reported 75.4% at maximum reasoning.
Second, Astra’s ARC-AGI-3 result measures the model together with OpenAI’s agent system. The company ran Astra with a Responses API harness that retains reasoning between turns and uses compaction to manage long contexts, and it has previously demonstrated that those settings substantially raise ARC-AGI-3 scores without changing the model. Epoch AI, which runs FrontierMath, has stated that OpenAI funded the benchmark’s development and holds exclusive access to part of it.
Third, Anthropic evaluated Fable 5.1 with production safeguards enabled. On tasks where those safeguards intervened, the model scored zero on OSWorld 2.0, which suppresses its published figure. OpenAI separately noted that its BenchCAD comparison used modified evaluation settings for the Claude results. Anyone tracking how AI benchmark scores are produced and reported should treat single-vendor charts as directional rather than decisive.
Cost per Task, Not Cost per Token
Both models list at $10 per million input tokens and $50 per million output tokens. The difference appears in how each company reduces effective spend.
Anthropic cut Fable 5.1 cache reads by 75%, to $0.25 per million tokens. Because cached context dominates agentic workloads, that single change lowers typical costs by roughly 25% relative to Fable 5, and context-heavy, tool-heavy work by up to about 45%. Cognition’s Walden Yan said the company moved its Opus 5 traffic in Devin to Fable 5.1 on launch day, noting it matched or edged out Fable 5 at a lower cost per task and that the new cache read pricing finally made a Fable-class model economical for code review.
OpenAI’s argument runs through token efficiency instead. President Greg Brockman put it plainly: “The price per task is what matters.” Astra uses fewer tokens on several evaluations and in partner tests, though OpenAI’s launch data is too thin to show whether those savings cover the premium over GPT-5.6 Sol, which remains at $5 and $30. If your team is still calibrating spend, our explainer on tokens as the working currency of generative AI sets out how these rates translate into monthly bills.
| Model | Input per 1M | Output per 1M | Cache read per 1M | Context window |
|---|---|---|---|---|
| Claude Fable 5.1 | $10.00 | $50.00 | $0.25 | 1,000,000 |
| GPT-6 Astra | $10.00 | $50.00 | Not disclosed at launch | 1,050,000 |
| Claude Opus 5 | $5.00 | $25.00 | Standard Opus rate | 1,000,000 |
| GPT-5.6 Sol | $5.00 | $30.00 | Standard rate | 1,050,000 |
How Each Model Handles Long Tasks
Fable 5.1: Endurance and Self-Verification
Anthropic’s pitch centres on work that runs for hours without a human watching. Ramp’s Dwight Temple described a single unattended 38-hour machine learning run that diagnosed a prior result as a label artifact, corrected it, launched six parallel overnight experiments and returned with a result and next steps. Shopify’s Ben Lafferty reported workflows running long stretches without losing the thread, with the model keeping its own records and reprioritising as conditions changed.
The quality claim is about root causes rather than patches. Millennium’s Damien recounted a crash occurring roughly once in a million runs that nobody on the team had explained in four to five years. Fable 5.1 disassembled an external vendor library, matched it against the core dump and traced the fault to a bug in that library. Effort levels give teams a lever here: the model defaults to High effort in Claude Code and Medium in Claude Cowork and on claude.ai, and at Low or Medium it matches or beats Fable 5 at considerably lower cost. Teams weighing this against cheaper Anthropic tiers can consult our breakdown of Claude Sonnet 5 against the rest of the Anthropic lineup.
Astra: Memory Across Context Windows
Astra’s practical gain for developers may matter more than its benchmark spread. Codex previously relied on compaction, summarising earlier work to free context, which discards precisely the detail an agent needs later: why a fix failed, which tests ran, which small requirement arrived at the start. Astra instead keeps notes across context windows and searches earlier messages and tool output. The feature sits behind a config.toml setting for now, and OpenAI says it becomes the default within weeks.
Astra can also ask a question without halting work that does not depend on the answer, which removes a common stall in coding agents. OpenAI demonstrated the model operating KiCad, Excel, Blender and Power BI, alongside browser form entry and website QA, and reported that Astra with the new Codex harness completed Mind2Web tasks 1.9 times faster than the Sol-based setup while cutting average desktop task time from about 75 minutes to 40.
Safety Behaviour Changes What Ships
Both launches tightened cybersecurity handling, and both now route sensitive work away from the default model.
OpenAI says Astra crossed the Critical cybersecurity threshold in its Preparedness Framework. In internal tests the model developed exploits for hardened browsers and operating systems and found two previously unknown vulnerabilities during evaluation against recent V8 bugs. The publicly available version refuses advanced cybersecurity work including exploit discovery, and vetted defenders receive less restricted access through Daybreak Blue. For API users, a cybersecurity safety check stops a task outright rather than pausing for approval. OpenAI’s Mia Glaese warned that users outside trusted-access programmes may hit slowdowns, pauses or blocks during cybersecurity work and occasionally during unrelated work.
Anthropic went the other direction on precision. Fable 5.1 can now be used to identify software vulnerabilities, though not to develop exploits for them, and the company says Claude Code users should expect around 60% fewer interventions per session from cyber safeguards. Penetration testing, exploit generation and binary-based vulnerability scanning still redirect to Opus models. Biology safeguards fire 85% less often for benign elementary biology and medical questions, while life sciences research and development routes to Mythos 5.1 through a verification programme built with the US government. The history behind those restrictions is covered in our report on the US government’s reversal of its Claude Fable 5 export ban.
Which One Should Your Team Pick?
| If your priority is… | Choose | Reason |
|---|---|---|
| Multi-hour unattended agent runs | Fable 5.1 | Strongest partner evidence for sustained autonomy and self-verification |
| Agentic coding in a terminal | Fable 5.1 | Leads Terminal-Bench 4.0 and CursorBench 3.2.0 |
| Scientific research workflows | GPT-6 Astra | 64.6% on Terminal-Bench Science against 52.6% |
| Abstract reasoning and hard mathematics | GPT-6 Astra | ARC-AGI-3 and FrontierMath Tier 4 results, with harness caveats |
| Screenshots or CAD views into code | GPT-6 Astra | 95.9% on BenchCAD Vision2Code |
| Documents, decks and business workflows | Fable 5.1 | Top GDPval-AA v2 and AutomationBench scores |
| Lowest bill on cache-heavy agents | Fable 5.1 | $0.25 per million cache reads |
| Defensive cybersecurity at scale | Depends on programme access | Mythos 5.1 through CVP, or Astra through Daybreak Blue |
Teams running both are the most common pattern so far. Anthropic’s model is the safer default for work that has to finish correctly and unsupervised; OpenAI’s is the stronger pick for research-grade reasoning and cross-application desktop breadth. Neither margin is wide enough to justify rebuilding a working stack, and both are positioned at the top of a market where the lead between the major AI labs keeps changing hands.
What to Watch Next
Anthropic says it is bringing Fable 5.1’s improvements to the rest of the Claude family, which would put much of this capability at Opus and Sonnet prices. OpenAI has not announced Luna, Terra or Sol variants for GPT-6; the lineup is currently Astra and Astra Pro only, leaving a gap at the cheaper end that GPT-5.6 still fills. Independent leaderboard results over the coming weeks will matter more than either launch chart, since both companies published scores using their own harnesses and their own readings of the other’s model.
This article covers model performance and pricing as published in September 2026. Benchmark results, prices and availability change frequently; verify current figures with the vendors before making procurement decisions.
If you are interested in this topic, we suggest you check our articles:
- Claude Mythos Guide: Anthropic’s Cybersecurity AI Model Explained
- Why the US Government Reversed Its Claude Fable 5 Ban
- Claude Sonnet 5 vs Anthropic’s Other Models Compared
- xAI vs OpenAI vs Anthropic: Which AI Lab Wins in 2026?
- Tokens Explained: The Currency of Generative AI
Sources: Anthropic — Introducing Claude Fable 5.1 and Claude Mythos 5.1, The New Stack — OpenAI launches GPT-6 Astra, OpenAI — GPT-6 Astra System Card, Claude Platform Docs — Claude Fable 5.1, MarkTechPost — Anthropic Releases Claude Fable 5.1 and Claude Mythos 5.1, SD Times — Anthropic releases Claude Fable 5.1 and Mythos 5.1, Wikipedia — GPT-6 Astra
Written by Alius Noreika
