GPT-6 Astra and Claude Fable 5.1 tie at the top and cost the same per token. One of them cost 59% less to run.
GPT-6 Astra (52.81) and Claude Fable 5.1 (53.37) tie on Artificial Analysis v4.3 at the same $10 / $50, yet Astra cost 59% less to run. Where Fable wins.
GTM Architect & Growth Operator · Now · 16 September 2026
TL;DR · Key insights
- On Artificial Analysis's Intelligence Index v4.3, Claude Fable 5.1 scores 53.37 and GPT-6 Astra 52.81. Artificial Analysis describes it as Astra tying Fable 5.1 for the lead.
- Both list at $10 / $50 per million tokens. Running the full index cost $13,129 on Fable 5.1 and $5,324 on Astra, because Fable 5.1 wrote 190M tokens to Astra's 60M.
- Fable 5.1 wins on two prices: cached input at $0.25 against Astra's $1, and prompts above 272K tokens, where Astra's whole request moves to $20 / $75 and Fable 5.1 keeps its rate.
- On coding agents they tie at 62. Codex with Astra costs $7.47 per task and takes 29.4 minutes; Claude Code with Fable 5.1 costs $12.39, takes 34.8 minutes and wins two of the three component benchmarks.
At the top of Artificial Analysis’s Intelligence Index v4.3 sit two models with the same price list, $10 per million input tokens and $50 output, and 0.56 points between them.
On paper there is nothing to choose. On a bill there is a lot.
| Info | GPT-6 Astra | Claude Fable 5.1 |
|---|---|---|
| Input / output per 1M | $10 / $50 | $10 / $50 |
| Cached input per 1M | $1.00 | $0.25 |
| Above 272K input tokens, in / out | $20 / $75 | $10 / $50 |
| Intelligence Index v4.3 | 52.81 | 53.37 |
| Coding Agent Index, own harness | 62 (Codex) | 62 (Claude Code) |
| Cost per coding task | $7.47 | $12.39 |
| Minutes per coding task | 29.4 | 34.8 |
| Cost to run the full index | $5,324 | $13,129 |
| Tokens generated on the index | 60M | 190M |
| Output speed, tokens/s | about 53 | about 65 |
| Knowledge cutoff | Apr 2026 | Jun 2026 |
Prices from OpenAI's GPT-6 Astra model page and Anthropic's pricing page; cutoffs from both vendors' model pages. Index to two decimals from Artificial Analysis's v4.3 chart data; run cost, tokens and speed from its model pages; coding figures from its Claude Code vs Codex comparison. Read on 16 September 2026.
A tie on both indexes
Artificial Analysis scores Fable 5.1 at 53.37 and Astra at 52.81 on v4.3, and describes it as Astra tying Fable 5.1 for the lead. Both display as 53 on its model pages.
On the Coding Agent Index they tie again at 62, each in its own vendor’s harness. The components split. Codex with Astra wins DeepSWE, the long-horizon repository tasks, 68% to 64%. Claude Code with Fable 5.1 wins Terminal-Bench 4.0, 58% to 56%, and SWE-Atlas-QnA, 65% to 62%.
Same price list, very different bills
Both charge $10 per million input tokens and $50 for output. What separates them is how much each writes.
Artificial Analysis counted 190M tokens from Fable 5.1 across its index and 60M from Astra, and spent $13,129 and $5,324 running the two. On coding agents, Codex with Astra cost $7.47 per task and Claude Code with Fable 5.1 $12.39, and Astra finished sooner as well, 29.4 minutes against 34.8.
That makes Astra the cheaper model on a typical agentic workload, even though no line on its price list is lower.
A tie on score and a tie on price per token still left a 2.5 times difference in what a full run cost. Output volume decides between these two.
Where Fable 5.1 is cheaper
Two lines on the price lists favour Fable 5.1.
Cached input. Fable 5.1 reads cache at $0.25 per million, a quarter of Astra’s $1. Suppose Fable 5.1 writes three times Astra’s output on your task, as it did across the index. Then its cache discount only wins when cached input is more than about 130 times Astra’s output, which means a very cache-heavy loop with short answers.
Prompts above 272K tokens. Once a prompt crosses 272K, OpenAI bills the whole Astra request at $20 input, $2 cached and $75 output. Anthropic keeps Fable 5.1 at one rate across its 1M window. A 300K-token prompt with a 5K answer costs $6.38 on Astra and $3.25 on Fable 5.1, which I worked through in the 272K pricing cliff.
What else separates them
- Speed. Fable 5.1 generates faster, about 65 tokens per second against Astra’s 53. Astra still finishes coding tasks sooner because it writes less.
- Knowledge. OpenAI lists Astra’s knowledge cutoff as 30 April 2026. Anthropic lists Fable 5.1’s reliable cutoff as June 2026.
- Refusals in life sciences. OpenAI’s launch table left Fable 5 and 5.1 out of three life sciences benchmarks because they refuse most of the questions. If your work sits near medicine or biology, test that first.
- Monitorability. OpenAI says Astra’s written reasoning is harder to monitor than GPT-5.6 Sol’s.
Which one I would use
| The work | Model | Why |
|---|---|---|
| Long agentic runs, prompts under 272K | GPT-6 Astra | Same score for well under half the cost of a full run. |
| Long-horizon repository changes | GPT-6 Astra | Wins DeepSWE in Codex, 68% to 64%. |
| Prompts that regularly cross 272K tokens | Claude Fable 5.1 | One rate across 1M tokens, while Astra's whole request doubles on input. |
| Very cache-heavy loops with short answers | Claude Fable 5.1 | Cached input at a quarter of Astra's rate. |
| Terminal work and questions about a codebase | Claude Fable 5.1 | Wins Terminal-Bench 4.0 and SWE-Atlas-QnA in Claude Code, narrowly. |
Level on capability. Output volume, prompt length and cache share decide the bill.
For the step below both inside Anthropic’s lineup, see GPT-6 Astra against Claude Opus 5 and Opus 5 against Fable 5.1. For the harnesses these scores were measured in, Claude Code against Codex.
What would change my mind
A same-harness run. Every coding number here pairs a model with its own vendor’s harness.
A change to either cache price or to OpenAI’s 272K rule. Those are the two lines where Fable 5.1 wins today.
The next index version. A 0.56-point gap can flip on a reweighting.
Questions people asked
Is GPT-6 Astra better than Claude Fable 5.1?
They are level. Artificial Analysis’s Intelligence Index v4.3 scores Claude Fable 5.1 at 53.37 and GPT-6 Astra at 52.81, which Artificial Analysis describes as a tie for the lead, and both score 62 on its Coding Agent Index. Within that tie, Claude Code with Fable 5.1 wins Terminal-Bench 4.0 and SWE-Atlas-QnA, and Codex with Astra wins DeepSWE, the long-horizon repository tasks.
Which is cheaper, GPT-6 Astra or Claude Fable 5.1?
They list at the same $10 per million input tokens and $50 output. Astra is cheaper on a full workload: Artificial Analysis spent $5,324 running its index on Astra against $13,129 on Fable 5.1, and measured $7.47 per coding task against $12.39. Fable 5.1 is cheaper on cached input, $0.25 against $1, and on prompts above 272K input tokens, where Astra’s whole request is billed at $20 input and $75 output.
Which is faster, GPT-6 Astra or Claude Fable 5.1?
Fable 5.1 generates faster, at about 65 output tokens per second against Astra’s 53 on Artificial Analysis’s measurements. Astra usually finishes first because it writes far less: on coding agents Codex with Astra averaged 29.4 minutes per task against 34.8 for Claude Code with Fable 5.1.
Should I use GPT-6 Astra or Claude Fable 5.1?
GPT-6 Astra for long agentic runs with prompts under 272K tokens, where it matches Fable 5.1 on score for well under half the cost of a full run. Claude Fable 5.1 for prompts above 272K tokens, for very cache-heavy loops with short answers, and for teams working in Claude Code.
If you are choosing a frontier model for long-running agents, the contact page is the fastest route to me.
