Claude Fable 5.1 vs GPT-6 Astra for coding

Claude Fable 5.1 and GPT-6 Astra, released two days apart in September 2026, cost the same per token ($10 per million input, $50 per million output), and Astra comes out slightly ahead for coding. In OpenAI's launch table Astra leads Fable 5.1 on Terminal-Bench 4.0 (57.9% to 55.8%), DeepSWE (74.1% to 67.4%) and FrontierCode (53.3% to 50.9%), while Fable 5.1 leads on Humanity's Last Exam with tools (65.0% to 57.2%) and Artificial Analysis's Intelligence Index (65.7 to 61.2). Independent runs are close: Vals AI's Terminal-Bench 4.0 run has Astra at 59.6% and Fable 5.1 at 58.1%, with Astra's tasks costing $9.58 against $17.18 and finishing in about 36 minutes against an hour. People's votes on LMArena's WebDev leaderboard put Astra second and Fable 5.1 fifth. Both trail Claude Opus 5.5 on those two independent rankings.

Key takeaways

• Same list price, different extras: Fable 5.1 has cheaper cache reads ($0.25 to $1) and no long-prompt surcharge; Astra has a fast mode and costs more on prompts over 272,000 tokens.
• Coding: Astra leads on most coding benchmarks, by up to 7 points, and was faster and cheaper per task on Vals AI's run.
• Reasoning and knowledge work: Fable 5.1 leads on Humanity's Last Exam and on Anthropic's GDPval-AA results.
• Access: Astra is included in ChatGPT Plus and up and in Codex. Fable 5.1 is usage credits on Claude Pro and up to half the weekly limits on Max.

Sources: OpenAI's GPT-6 Astra post and model page; Anthropic's Fable 5.1 announcement, docs and Opus 5.5 post; Vals AI and LMArena, opened October 27, 2026.

How much do Fable 5.1 and GPT-6 Astra cost?

Per million tokens on each lab's API:

• Input and output: $10 and $50 for both.
• Cached input: $0.25 on Fable 5.1 and $1 on Astra. Cache writes are $12.50 on both (Fable also has a 1-hour cache at $20).
• Long prompts: Fable bills its whole 1M-token window at the standard rate. Astra charges twice the input and cache rates and 1.5 times the output rate, for the whole request, when a prompt passes 272,000 tokens.
• Batch: half price on both.
• Fast mode: Astra's costs twice the standard price for up to twice the speed. Fable has none.

For long agent sessions with large cached prompts, Fable's pricing is the cheaper of the two per token. Measured per task, Astra came out cheaper: on Vals AI's Terminal-Bench 4.0 run, $9.58 against $17.18, because Astra used fewer tokens and less time. More on Fable's rates in Claude Fable pricing.

How do they compare on coding benchmarks?

OpenAI's launch table (maximum score at any effort, Fable 5.1 numbers as reported or reproduced by OpenAI):

• Terminal-Bench 4.0: Astra 57.9%, Fable 5.1 55.8%.
• DeepSWE v1.1: 74.1% against 67.4%.
• FrontierCode 1.1, Main: 53.3% against 50.9%. Extended: 64.5% against 63.6%.
• Internal database migration tasks: 63.9% against 57.8%.

Independent results:

• Vals AI, Terminal-Bench 4.0 (one shared harness, updated October 26): Astra 59.60% at $9.58 and about 36 minutes a task; Fable 5.1 58.08% at $17.18 and about an hour.
• LMArena WebDev (people voting on web builds, October 27): Astra at max effort is second with 1774, Fable 5.1 fifth with 1745.

Neither is first on those two: Claude Opus 5.5 leads both, at 65.15% and 1809. See the best AI model for coding.

Where does Fable 5.1 lead?

• Humanity's Last Exam, with tools: 65.0% against 57.2% in OpenAI's own table.
• Artificial Analysis Intelligence Index v4.1.1: 65.7 against 61.2, also from OpenAI's table.
• Knowledge work: Anthropic's Opus 5.5 post puts Fable 5.1 at 1735 on GDPval-AA v2.1 and Astra at 1542.

Astra leads on most of the rest of OpenAI's table, including computer use, math (FrontierMath Tier 4: 97.6% to 87.8%), science (Terminal-Bench Science: 64.6% to 52.6%) and business workflows (AutomationBench: 41.4% to 31.4%).

How are they different to use?

• Context: Fable 5.1 takes 1M tokens with 128,000 output. Astra takes 1.05M (922,000 of it input) with 128,000 output, and its knowledge cutoff is April 30, 2026.
• Effort: both are reasoning models with adjustable effort. Astra offers low to max; Fable 5.1 can change effort mid-conversation on the API (a beta).
• Data: Astra supports zero data retention for eligible API customers. Fable 5.1 carries 30-day retention and isn't available under zero data retention unless Anthropic authorizes it.
• Security work: OpenAI rates Astra at the Critical level for cybersecurity under its Preparedness Framework; the shipped model helps with secure code review and patching but refuses proof-of-concept exploits. Fable 5.1 can help find vulnerabilities but not write exploits, and Mythos 5.1, the same model with fewer safeguards, is limited to verified organizations.

Where can you use each?

• GPT-6 Astra: ChatGPT Plus, Pro, Business and Enterprise, inside the subscription's allowance (Enterprise admins turn it on), plus Codex, the OpenAI API, Azure and Amazon Bedrock. Pro, Business and Enterprise also get Astra Pro. Codex usage on Astra is covered in Codex pricing.
• Claude Fable 5.1: Claude Pro with usage credits, Max and premium seats up to 50% of their weekly limits, Claude Code (version 2.1.255 or later), the API, Bedrock, Google Cloud and Microsoft Foundry. Not on Claude's free plan.

Can you use both together?

Each lab's own apps run only its own models. GitHub Copilot offers both. In Poly (usepoly.co), Fable 5.1 runs in Claude Code and GPT-6 Astra in Codex, both on paid plans, and each person in a shared room picks the model for their own turns. A team can ask Astra to review what Fable built, and Poly keeps a record of which model wrote each reply. See cross-model code review and Claude vs ChatGPT for coding.