Is GPT-6 Astra Better Than Fable?
GPT-6 Astra is not better than Claude Fable 5.1 on the independent scoreboard that actually ranks general intelligence.
Artificial Analysis put Astra at 61 on Intelligence Index v4.1.1 on 3 September 2026, and Fable 5.1 at 66. OpenAI's own launch table still prints that 61.2 vs 65.7 row, then leads with "the world's most intelligent model."
We ran an OpenRouter ping for openai/gpt-6-astra at 14:40 UTC on 4 September 2026. The gateway returned HTTP 400: not a valid model ID. Fable 5.1 was already on the same list. Our setup is that gateway, not a private OpenAI key.
GPT-6 Astra vs Fable 5.1: which one should you keep?
Keep Fable 5.1 this morning; Astra wins computer-use and math on vendor rows, not the independent index.
The Claude Fable 5.1 review already owns the 5.1 vs 5 vs Opus 5 invoice. This table is the cross-lab split.
| Door | Callable on OpenRouter 2026-09-04 | AA Intelligence Index | AA Coding Agent Index | List price (in/out) | Cache read |
|---|---|---|---|---|---|
| Claude Fable 5.1 | Yes (anthropic/claude-fable-5.1) | 66 | 70 | $10 / $50 | $0.25 |
| GPT-6 Astra | No (HTTP 400) | 61 | 67 | $10 / $50 | $1.00 |
| GPT-5.6 Sol | Yes (openai/gpt-5.6-sol) | 61 | 65 | $2 / $10 on OpenRouter; AA still cites $4 / $20 list | $0.20 / 90% off |
Sources: Artificial Analysis Astra and Fable 5.1, 3 September and 1 September 2026; OpenAI model card for Astra list price; our test of the callable column is the OpenRouter ping.
Astra is the computer-use and FrontierMath story. Fable 5.1 is the intelligence-index and cache-read story.
Sol is the model you can already call at a fraction of the $10/$50 card. Treat "better" as a workload, not a headline.
Can you even call GPT-6 Astra today?
Not from OpenRouter. openai/gpt-6-astra returned HTTP 400 at 14:40 UTC on 2026-09-04.
OpenAI's launch post says a limited set of organizations got it on 3 September, with ChatGPT Plus, Pro, Business, Enterprise, the API, Azure, and Bedrock following "over the coming days." Free-tier access is unannounced. Enterprise workspaces ship with Astra off.
anthropic/claude-fable-5.1 and openai/gpt-5.6-sol were both callable on OpenRouter the same afternoon.
If your stack already sits behind a gateway, that 400 is the only number that matters until the ID resolves.
curl -sS https://openrouter.ai/api/v1/chat/completions \
-H "Authorization: Bearer $OPENROUTER_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model":"openai/gpt-6-astra","messages":[{"role":"user","content":"PONG"}],"max_tokens":16}'
# HTTP 400 {"error":{"message":"openai/gpt-6-astra is not a valid model ID","code":400}}
# probed 2026-09-04T14:40:37ZRe-run that ping before you rewrite a router. A blog post dated 4 September will go stale the hour the ID lands.
Do independent benches agree Astra is smarter?
No. Artificial Analysis puts Fable 5.1 at 66 and Astra at 61 on Intelligence Index v4.1.1.
OpenAI still wrote "world's most intelligent" above a table that already contains that loss. The coding story is the same shape: OpenAI calls Astra the best software-engineering model to date; AA's Coding Agent Index has Fable 5.1 at 70 and Astra at 67.
The ARC-AGI-3 headline is a runner split, not an AGI certificate. OpenAI printed 99.9%. ARC Prize ran the same model on its provider-neutral Standard runner and got 62.7% at max effort for $26,098.
The 99.9% run used a Provider Adapter that keeps OpenAI's hidden reasoning state between turns, at high effort, for $18,817. Claude Opus 5 sits at 30.2% on that Standard board. GPT-5.6 Sol sits at 7.8%. Astra doubled the prior Standard SOTA. It did not saturate the benchmark under the runner every other lab can share.
| Bench | GPT-6 Astra | Claude Fable 5.1 | Who ran it |
|---|---|---|---|
| AA Intelligence Index v4.1.1 | 61 | 66 | Artificial Analysis |
| AA Coding Agent Index | 67 | 70 | Artificial Analysis |
| Terminal-Bench 4.0 | 57.9% | 55.8% | OpenAI |
| DeepSWE v1.1 | 74.1% | 67.4% | OpenAI |
| AutomationBench | 41.4% | 31.4% | OpenAI |
| Terminal-Bench Science 0.1 | 64.6% | 52.6% | OpenAI |
| FrontierMath Tier 4 v2 | 97.6% | 87.8% | OpenAI |
| GPQA Diamond | 96.0% | 93.7% | OpenAI |
| Humanity's Last Exam (w/ tools) | 57.2% | 65.0% | OpenAI |
| ARC-AGI-3 Standard | 62.7% ($26,098) | not scored | ARC Prize |
| ARC-AGI-3 Provider Adapter | 99.9% ($18,817) | not scored | OpenAI / ARC Prize |
| Vals Index | 66.61% · $19.09/test | not fetched | Vals.ai, 4 September 2026 |
AA also spent $3,020.82 to run Astra through the Intelligence Index. Token-thrift versus Sol is real. It does not move the index past Fable 5.1.
Hallucination on AA-Omniscience dropped from 92% to 51% at max effort versus Sol. That is the cleanest independent win Astra has on knowledge reliability.
OpenAI's "most intelligent" line and ARC Prize's 62.7% Standard score cannot both be the whole story. Print both.
What does the $10/$50 invoice do vs Fable?
Same list price, worse cache reads: Astra $1.00 vs Fable 5.1 $0.25 per million.
The GPT-6 Astra model card lists $10 input, $50 output, $1.00 cached input, $12.50 cache writes, a 1,050,000-token window, 128,000 max output, and an April 30, 2026 knowledge cutoff. Prompts over 272k input tokens bill 2× on input and cache and 1.5× on output. Fast mode is 2×. Batch and Flex are 50%.
Fable 5.1 sits on the same $10/$50 card. The sibling already metered the cache-read cut to $0.25. Do not rebuild that probe here; cite the Fable 5.1 review and use the list numbers.
A 30-day job at 2.4 million input tokens and 380,000 output tokens per day, at $10/$50, is:
2,400,000 × 30 × $10 / 1,000,000 = $720 input
380,000 × 30 × $50 / 1,000,000 = $570 output
= $1,290That $1,290 is identical on Astra and Fable 5.1 before cache. Forty rereads of a 200,000-token cached prefix are not:
| Cache-read line | Rate | 40 × 200,000 tokens |
|---|---|---|
| Claude Fable 5.1 | $0.25 / MTok | $2.00 |
| GPT-6 Astra | $1.00 / MTok | $8.00 |
Four times the cache line, same output card.
For the rest of the inference cost model, that gap compounds on any agent that rereads a system prefix. AA's own write-up says Astra is about 75% more expensive per Intelligence Index task than Sol at max effort because the 2.5× price hike eats the ~10% token cut.
Sol remains the cheap OpenAI door.
Which real-world jobs is Astra actually for?
OpenAI reports Astra finishes OSWorld 2.0 at 72.6% in about 40 minutes, versus Sol at 65.7% in 75.
That is the computer-use claim, and it is OpenAI's number, not ours. ScreenSpot-Pro without tools is 92.7% against Sol's 76.9%. Agents' Last Exam is 59.3% against Sol's 53.6% and Fable 5's 48.7% (OpenAI did not print a Fable 5.1 row there).
Attributed demos from the launch post, not from our desk:
- Desktop and browser agents. OSWorld 2.0 partial, ~47% less wall time than Sol. Pair it with coding-agent tooling only after the API ID resolves.
- Long Codex sessions. OpenAI is shipping notes that survive compaction, so failed-fix context is searchable instead of summarized away.
- Game and spatial work. Playco says Astra cut manual fixes about 50% on Unity/Godot prototypes. That is Playco talking, in an OpenAI case study.
- Defensive cyber. Astra is the first OpenAI model rated Critical under the Preparedness Framework. The production build refuses proof-of-concept exploits. Advanced workflows sit behind Daybreak / Trusted Access.
- Math that saturates a board. FrontierMath Tier 4 v2 at 97.6% vs Fable 5.1 at 87.8%, vendor-run. Epoch AI funds part of that board; treat it as OpenAI-adjacent.
What it is not for, today: a drop-in model= swap on OpenRouter, a cheaper Fable, or a claim that 99.9% on ARC-AGI-3 means AGI.
Greg Brockman told reporters "Welcome to the AGI era." WIRED printed the quote. ARC Prize, which built the benchmark, is not claiming AGI.
Should you stay on Fable 5.1, wait for Astra, or stay on Sol?
Stay on Fable 5.1 until Astra's model ID resolves on your gateway and your work is computer-use or FrontierMath.
Route on workload, not on the launch blog.
| If this is true | Keep | Flip when |
|---|---|---|
| You need a model this morning on OpenRouter | Fable 5.1 (or Sol if the $10/$50 card is the problem) | openai/gpt-6-astra returns 200 on the ping above |
| Your eval is AA Intelligence Index / HLE-with-tools / cache-heavy agents | Fable 5.1 | Independent Index actually moves past 66, or your cache-read bill is not the line that matters |
| Your eval is OSWorld, ScreenSpot-Pro, AutomationBench, FrontierMath | Wait for Astra | The ID resolves and you can measure the job on your own traces |
| You are still on GPT-5.6 Sol and the work is structured, short, cheap | Sol (Sol Ultra / Codex if that is already the lane) | A computer-use or math job shows up that Sol's 7.8% Standard ARC and 61 Index cannot carry |
| You need offensive cyber PoCs | Neither public SKU | Daybreak / Trusted Access, with the refusals still on the public build |
Fable 5.1 is the default. Astra is the computer-use upgrade you cannot buy on the gateway we ran this ping against.
Sol is the OpenAI SKU that already answers. Re-run the curl, then move.
Frequently Asked Questions
Is GPT-6 Astra AGI?
No. Greg Brockman used "AGI era" on a press call. That is a remark, not a measured finding. ARC Prize scored Astra at 62.7% on the Standard ARC-AGI-3 runner and said it is not claiming AGI. OpenAI's 99.9% figure used a Provider Adapter that keeps hidden reasoning state between turns. A saturating adapter score is not the same test every other model sat.
Can free ChatGPT users get GPT-6 Astra?
Not according to the launch post. OpenAI named Plus, Pro, Business, and Enterprise, plus the API, Azure, and Bedrock, for the "coming days" wave. Free-tier access is unannounced. Enterprise admins get a toggle that defaults to off. Treat "today" as Daybreak / Trusted Access unless your workspace admin has flipped that switch.
When will GPT-6 Astra show up on OpenRouter?
Unknown. Our 4 September 2026 14:40 UTC ping returned HTTP 400 for openai/gpt-6-astra. Re-run the curl in the access section. Do not rewrite a router on a launch-blog promise. The moment the ID resolves, the callable-today column in the first table flips.
Why do some pages say 99.9% on ARC-AGI-3 and others say 62.7%?
Two runners. ARC Prize's Standard runner is provider-neutral and scored 62.7% at max effort ($26,098). The Provider Adapter preserves OpenAI's opaque reasoning state and scored 99.9% at high effort ($18,817). OpenAI led with the adapter number. Always print the pair.
Will GPT-6 Astra write proof-of-concept exploits for me?
The public build will refuse that class of request. OpenAI rated Astra Critical for cybersecurity under its Preparedness Framework and limited advanced exploit work to Daybreak / Trusted Access. ExploitBench 100% was measured without production safeguards. Defensive review and patching are the intended public cyber jobs.