AI Source-backed visual comparison 11-minute read
Claude Fable 5.1 Max vs GPT-6 Astra: Benchmarks & Pricing
Compare Claude Fable 5.1 and GPT-6 Astra on coding, reasoning, and API pricing. Explore source-linked charts, a cost calculator, and what Max really means.
The short version
- Astra leads on several published coding and science tests. Fable leads on other tough tests. There is no winner for every job.
- Their starting prices match, but Fable charges less to reuse saved input. Astra charges more for very long requests.
- Max is a setting that lets a model spend more effort on an answer. It is not a promise of better work.
Based on provider documentation and published results—not a Curio head-to-head test. How this comparison was made.
Explore the test results and estimate your bill. The model symbols are Curio illustrations.
Which wins at coding and reasoning?
GPT-6 Astra scores higher on the coding tests below. Claude Fable 5.1 leads on Humanity's Last Exam with tools and the Artificial Analysis index. The choice gets more interesting when you add cost: their starting token prices match, but the same-sized request can produce a different bill.
These are published launch results, not a Curio head-to-head test. Start with the tests closest to your work, then use the calculator to explore the bill.
Max lets a model spend more effort on an answer. Both models offer that setting. It is separate from Claude Max, the paid plan.
The lead changes with the test.
These are provider-reported results. OpenAI’s launch table reports each model’s best result at any effort level, using different setups—not both set to Max.
How this comparison was made: these charts use the launch-table results reviewed , not a live leaderboard or a Curio test. Each chart links its source.
AutomationBench
Higher is betterCan the model finish a job that takes several steps and tools? Astra’s reported score is about ten percentage points higher. Published result
Terminal-Bench 4
Higher is betterCan it solve tasks by typing computer commands? Astra’s reported score is slightly higher. Published result
DeepSWE v1.1
Higher is betterCan it solve software engineering tasks? Astra’s reported score is roughly seven percentage points higher. Published result
TB Science 0.1
Higher is betterCan it carry out science tasks with computer tools? Astra’s reported score is twelve percentage points higher. Published result
GPQA Diamond
Higher is betterCan it answer hard science questions at graduate level? Both score highly; Astra’s reported score is slightly higher. Published result
Humanity’s Last Exam · with tools
Higher is betterCan it answer hard expert questions across many subjects, using tools? Fable’s reported score is about eight percentage points higher. Published result
Artificial Analysis Intelligence Index
Index points · higher is betterThis combined score uses several tests (version 4.1.1). Fable leads by about five index points, not five percent more correct answers. Published result
Tools and time limits can change a score. All charts start at zero on a 0–100 scale; the last uses index points, not percentages. A ten-percentage-point gap means ten more out of a hundred on that test’s scale, not ten percent better at everything.
Context window: how much can they read?
The context window is how much material fits in a request. It is measured in tokens: small pieces of text, often parts of words. Astra has slightly more room; both models share the same output limit, which includes the answer and internal thinking.
Think of a bigger desk. It can hold more documents, but that does not tell you how well someone will find the important detail among them.
How much fits on the desk?
Input and reserved output must fit within the context window.
API pricing: why can the same job cost more?
The API is the service an app uses to send work to a model. Its per-request charges are separate from monthly chat plans.
Repeated work is where the bills can diverge. A cache lets the service reuse earlier input work—for example, a long style guide sent with every draft. Fable charges less to read that saved input. Astra also raises its rates for very long requests.
Try moving the calculator's reuse slider. It changes one part of the bill, not the cost of the new answer.
What would the same-sized job cost?
Change the example to see how input, reuse and output affect the estimated API bill.
At these example inputs, Fable costs six cents less.
| What gets billed | Fable 5.1 | GPT-6 Astra |
|---|---|---|
| New input | $10 | $10 |
| Output, including thinking | $50 | $50 |
| Reusing saved input · cache read | $0.25 | $1 |
| Saving input · cache write | $12.50 · 5 min | $12.50 · 30 min |
At regular prices, Astra charges four times as much to read the same number of saved tokens. That is one part of the bill, not a fourfold difference in the whole job.
Long requests: above 272,000 input tokens, Astra charges 2× for all input and saved-input work, and 1.5× for all output. Fable keeps regular prices up to its 1M context limit.
What this estimate leaves out: tool fees, storage, tax, speed upgrades, grouped-request discounts, and chat plans. It compares equal token counts, not identical text—the models can split the same words differently. For a like-sized comparison, this calculator caps input plus output at 1M tokens. Astra’s own context limit is slightly higher. Anthropic pricing · OpenAI pricing · How long Astra saves input
Reasoning and tools: how do they work?
A model can ask to search the web or run code. The app runs the tool and returns its result. What happens while it waits depends on the model and the app's support.
Fable's adaptive thinking is always on and can continue between tool calls. The thinking shown on screen is a summary; hiding it does not remove the internal work or its cost.
Astra can work on an independent step while a tool runs, and accept new instructions mid-answer. Both features need support from the app; choosing Astra alone does not turn them on.
What happens between the question and the answer?
This is a tool-use loop, not a picture of either model’s undisclosed internal design.
Plan
Choose a next stepWork out what is needed to answer the request.
Use a tool
Search or take an actionAsk the app to search, run code or read a file.
Read the result
Use what came backUse the tool’s result to choose the next step.
Check
Test the answerCheck the source or run the test, not just say it was checked.
It may write a long answer twice.
- Draft an answer internally
- Write the answer again
- Extra cost and waiting
Anthropic warns that Max can produce this extra work. It is a documented risk, not a measured failure rate.
Anthropic prompting guideA tool can run while other work continues.
Async tool calls allow this overlap when the app supports it. They do not establish a faster end-to-end result.
OpenAI async tool guideCreative writing: which one is worth reading?
A coding score cannot tell you which model writes a better opening. Give both the same source and ask them to explain it to someone new to the subject.
Then ask a reader to explain the idea back. A fluent draft that leaves them confused—or confidently repeating something false—has missed the point.
The watermark brief below is one way to try that comparison. It has not been run; there is no Curio writing winner to announce.
Where would you notice the difference?
Use these differences to choose a first test, not a permanent winner.
| The job | Fable 5.1 | GPT-6 Astra | What to check |
|---|---|---|---|
| The same long guide sent with every task | Costs less to reuse saved input. | Charges more for that same amount of saved input. | Check how often the service can actually reuse the input. |
| A huge document to read at once | Regular prices up to its 1M context limit. | Slightly more space; higher prices above 272K input tokens. | Ask whether the whole document is needed on every turn. |
| A search or tool that takes a while | Available tools depend on the app or service. | Can keep working while a tool runs, in apps that support it. | Check what your app supports, not just which model it uses. |
| An explanation someone will remember | No matched Curio writing result. | No matched Curio writing result. | Compare drafts from the same source. Can a reader explain the idea back? |
Explain a watermark made of ordinary words.
Give both models the same brief: turn Anthropic’s explanation into an opening, a short article, and a visual, using only that source.
The idea to explain
A watermark can hide in a pattern of ordinary word choices.
The facts to preserve
No hidden characters, personal tracking, or guaranteed detection.
Hide the model names while readers compare the drafts. Ask what they learned, then check it against the source. Keep the fixes, time and cost beside each version.
The explanation below shows the idea; it is not an output from this proposed test. How Claude text watermarking works
Ordinary words. A pattern you don’t see.
When several words fit, Anthropic’s watermark method uses a secret key to help choose among them. Across a passage, those choices leave a pattern a detector can check. There are no hidden characters.
It is a clue that Claude was involved, not proof. It does not identify the person who used it.
Why can’t one word give it away?
A single choice could happen by chance. A longer passage gives the detector more choices to check. Very short text may not give the detector enough clues.
Which model should you choose?
Pick one real task, not a vague contest to find the smartest model. A daily draft, a difficult coding repair and a source-heavy report need different things.
Keep the finished result beside the bill, waiting time and repairs it needed. A cheaper answer that takes an hour to fix may be the more expensive choice for you. The useful winner is the one that helps finish your job.
Speed, training, privacy, and access
Policies checked . Access guidance checked .
Which model is faster?+
This guide does not report a matched speed test. Compare the complete job with the same brief and tools, including waiting and repairs. Astra’s support for overlapping work alone cannot settle the result.
Astra tool handlingHow were they trained?+
Anthropic describes public, private and AI-made training data, followed by further training to shape Fable’s behavior. Neither provider’s report gives the full internal design or parameter count—the number of learned settings. There is no supported size leaderboard to show here.
Fable training · Astra training overviewWhere does the data go?+
Anthropic normally keeps Fable request data for safety checks for 30 days, with exceptions for some business customers. OpenAI does not use API data for training by default; logs used to check for abuse normally stay for up to 30 days. The exact rules depend on the service and account settings. A monthly chat plan may have different rules.
Fable retention · OpenAI data controlsImages, audio, and video+
Both can read text and images and answer with text. Astra can also use a separate tool to make images. Neither model page lists built-in audio or video support. An app may add separate tools for those tasks.
Can everyone use them?+
Fable 5.1 is available through the API and paid Claude plans. Pro and standard Team seats use paid usage credits; Max includes limited Fable usage. Astra access depends on the app and account. OpenAI still lists restrictions for its early Enterprise rollout. ChatGPT access does not automatically grant API access. Check your model picker and the provider’s access page before paying.
Fable plans and access · Astra account and workspace accessWhat can make work harder?+
Fable’s guide warns about dense writing and rewriting a whole file when a small edit would do. Astra’s warns about asking too many questions and running more tests than needed. Fable’s and Astra’s guides describe possible problems, not how often the models failed in a shared test.
Fable guidance · Astra guidanceWhich writes better in another language?+
These sources do not settle which model writes best in each language. Compare translations and original writing with readers of that language. An English coding score cannot tell you which Spanish explanation feels clear and natural.
Fable language evaluations · Astra evaluationsSources
- OpenAI — Introducing GPT-6 Astra, comparison table and evaluation notes
- OpenAI — GPT-6 Astra model specifications
- Anthropic — Claude Fable 5.1 overview
- OpenAI — API pricing
- Anthropic — API pricing
- OpenAI — Prompt caching
- Anthropic — Reasoning effort
- Anthropic — Thinking
- OpenAI — Using GPT-6 Astra
- Anthropic — Tools overview
- Anthropic — Fable 5.1 prompting guide
- Anthropic — How Claude's text watermark works
- Anthropic — Fable and Mythos 5.1 system card
- Anthropic — Fable access and safeguards
- Anthropic — What is the Max plan?
- Anthropic — What's new in Fable 5.1
- OpenAI — API data controls
- Anthropic — Fable models on your plan
- OpenAI — Workspace model availability