GPT-6 Astra or Claude Fable 5.1: which to choose in 2026

Two flagship models launched in early September 2026: Claude Fable 5.1 from Anthropic on the first, and GPT-6 Astra from OpenAI on the third. The companies call them their best models for coding and complex work at the time of release. Both have the same base API price: $10 per million input tokens and $50 per million output tokens (a token is a unit of text that you pay for in the API).

When choosing, consider the tool you already work in, your tasks, and the costs of cache reads and long context. The models are included differently in subscriptions, so your monthly bill depends on both plan terms and token usage. This comparison uses official OpenAI and Anthropic pages as of September 20, 2026.

For information on working with the service, see the Ailoot documentation.

GPT-6 Astra: release and key capabilities#

OpenAI calls Astra its best model for computer control: filling out forms, working in a CRM, researching in a browser, and checking a website after it has been built. According to OpenAI's own measurements on the OSWorld 2.0 benchmark, Astra scores 72.6% versus 65.7% for the previous GPT-5.6 Sol and takes about 47% less time per task. Paired with the updated Codex, task completion on the Mind2Web benchmark was 1.9 times faster. These are the manufacturer's figures on its own configurations: a point of reference, not an independent measurement.

In Codex, Astra can keep notes between context windows instead of compressing a long session into a single summary. Earlier windows remain available for search. During a long debugging session, the model is less likely to lose track of why a previous change failed. This is still an experimental feature, and OpenAI promises to make it standard for Astra in the coming weeks. You can enable it in the Codex configuration by adding this to the ~/.codex/config.toml file:

[features.context_management]
experimental_mode = true

According to the official Codex help, you need to sign in with a ChatGPT account on the Plus, Pro, or Pro Lite plan. You can check whether the feature is enabled with the codex features command, which shows the flags and their values.

According to OpenAI's help page as of 20.09.2026, Astra is included in ChatGPT Plus for $20 per month and above, while the advanced GPT-6 Astra Pro chat mode is available only on Pro plans (5x for $100 or 20x for $200), Business, and Enterprise. In Enterprise, an administrator enables access. Since September 10, 2026, OpenAI has temporarily paused new subscriptions and upgrades to Pro for $200; existing $200 Pro subscriptions continue to work. Pro for $100 remains available.

• Astra asks clarifying questions. When the result depends on the answer, the model asks first instead of guessing. In Codex, questions can be asynchronous: work that does not depend on the answer continues.
• Gradual access rollout. Astra is OpenAI's first model to reach the Critical threshold for cybersecurity in the company's risk assessment system. As a result, access is being opened gradually, and the model refuses advanced vulnerability-related requests. The announcement specifically names advanced cybersecurity tasks among the restrictions; OpenAI does not report restrictions on ordinary code or documents. This explains why the model is not appearing for all users at once.

Claude Fable 5.1: cache reads cost four times less than with Astra#

Anthropic positions Fable 5.1 for long sessions where an agent works independently: tasks across an entire code project, multi-step research, documents, spreadsheets, and slides. For most everyday tasks, Anthropic recommends starting with Opus 5 and using Fable 5.1 for complex reasoning and long autonomous work. In Claude Code, it is set to a high reasoning effort level (high effort); in claude.ai and Claude Cowork, it is set to medium.

Fable 5.1 cache reads cost $0.25 per million tokens, compared with $1 for Astra—four times less. This is a comparison of cache read rates, not the total cost of a session: it does not follow that each repeated read of project files is charged at this rate. Anthropic estimates that typical tasks have become about 25% cheaper, and tasks involving active agent work up to 45% cheaper. This is the manufacturer's estimate; the cache read price is listed in the documentation.

Fable 5.1 has access restrictions, safeguards, and data retention terms:

• Mythos 5.1 is the same model, but it is not for everyone. Fable 5.1 and Mythos 5.1 are the same model with different levels of safeguards. Mythos is available only to participants in Anthropic's trusted programs. Fable 5.1, the full standard version, is available to ordinary users.
• Safeguards sometimes redirect a request. The system routes tasks in high-risk areas (offensive cybersecurity, biology) to a backup model from the Opus family, while searching for vulnerabilities in source code is now permitted on full Fable 5.1. Anthropic reports about 60% fewer cyber-safeguard interventions in Claude Code than with Fable 5. Redirection applies to high-risk areas, but for specific topics the response may come from another model.
• Data retention. For Fable models, Anthropic retains data for 30 days by default. Zero data retention is currently available to certain enterprise customers by arrangement, and starting this fall Anthropic promises a mode in which data is stored in the customer's own infrastructure. If you give the model sensitive company documents, check this point before starting work.

The result depends not only on the model, but also on the context you give it: project instructions, knowledge structure, and agent work rules.

Comparison across the same criteria#

Benchmarks show different results. On its Terminal-Bench Science benchmark, OpenAI reports an Astra result of 64.6% versus 52.6% for Fable 5.1, with an estimated API cost about 31% lower. At the same time, the OpenAI announcement's table includes an independent index where Fable 5.1 leads: 65.7 versus 61.2 for Astra. Anthropic publishes its own charts, where Fable 5.1 is noticeably stronger than Fable 5. When reading manufacturers' measurements, keep in mind that they choose the configurations themselves. The figures presented show no clear overall leader, and your task may differ from those in the benchmarks.

How do I select Claude Fable 5.1 or GPT-6 Astra for a task?#

• You write code every day in Claude Code. Select Fable 5.1 through /model and check it on your own tasks: Anthropic recommends it specifically for long agent sessions. When comparing costs, consider the cache read rate, but do not assume that every repeated project read is automatically cached. Your plan matters too: on Max (from $100), Fable is included in the weekly limit and can use up to half of it, while on Pro for $20 it works only with credits at API prices—that is, beyond the subscription.
• You write code in Codex. Astra is included in the ChatGPT Plus subscription limits for $20, with no extra charge for the model itself. The new feature for keeping notes between context windows is useful for long debugging sessions; how to enable it is shown above in the section about Astra.
• Browser tasks: fill out a form, collect data from websites, work in someone else's interface. This is Astra's main stated strength—computer control. If your automation works with websites, start testing with it.
• Documents, spreadsheets, and slides using a branded template. Both models claim to work with documents; OpenAI specifically emphasizes following existing templates. Give both models the same template of your own and compare the results.
• Huge input data. If your requests regularly exceed 272 thousand tokens (for example, an entire large document database), those requests cost Astra considerably more than the base price. Calculate the cost in advance for your volume.
• The smallest budget. Base prices are the same. The actual difference depends on how many tokens the model uses for your typical task. You can see this only by measuring it yourself.
• Change nothing. GPT-5.6 Sol and Claude Opus 5 remain usable models, and they are sufficient for most everyday tasks. Switching models makes sense when their capabilities are no longer enough for your tasks, not just because a press release came out.

How to test your choice on your project#

1. Choose one task from your work: a new feature in a project, a report, or data analysis. Pick something you actually need, not a toy example.
2. Do it with the first of the two models: Astra in Codex or Fable 5.1 in Claude Code, depending on what is available to you. Record how much time it took and how many tokens were used—both programs show usage statistics.
3. Give the same task, with the same description and files, to the second model in its tool. Don't have a second subscription? Compare the new model with the one you currently work with.
4. Compare three things: usage, time, and, most importantly, how many edits you made before the result worked. A cheap answer that had to be redone three times usually costs more than an expensive one.
5. Write down which scenarios each model handled better, and assign them accordingly. Repeat the test on a new task in a month—the models update quickly.

You don't have to maintain two subscriptions just to run these checks: start with the tool you already have, and consider the second only when your current one is no longer enough.

Frequently asked questions#

Are Fable 5.1 and Mythos 5.1 different models?

They are the same model. Mythos 5.1 is available only to participants in Anthropic's trusted programs for sensitive areas such as cybersecurity. Fable 5.1 is available to Claude Code and claude.ai users.

Is Fable 5.1 included in a Claude Pro subscription for $20?

It is not included in the plan limits. On Pro, it works with credits at API prices, meaning it costs extra beyond the subscription. It is included in the weekly limit only on Max (from $100 per month), where it can use up to half of the weekly allowance. You can check your plan in Settings, in the Usage section.

Do I need to switch to the new model right away?

There's no rush. See what has changed for your tasks: if you work in Claude Code, Fable 5.1 is selected through /model (with credits on Pro, within the limit on Max); if you work in Codex, Astra is already included in Plus. Change your usual workflow after testing it on your task, not on announcement day.

What will happen to GPT-5.6 Sol and Fable 5?

Both remain available. GPT-5.6 Sol is still suitable for most everyday tasks, while Fable 5.1 kept Fable 5's input and output prices and made only cache reads cheaper.

Is it true that Astra is dangerous?

OpenAI has, for the first time, marked one of its models as having reached the Critical threshold for cybersecurity in its risk assessment system. The practical consequence for ordinary users is a gradual access rollout and refusals on requests about creating exploits; the announcement says nothing about restrictions on ordinary code, documents, or browser tasks.

What next#

Open the tool you already work in and test the new model on one real task using the steps in “How to test your choice on your project.” If you are unsure between the ecosystems as a whole, see our comparisons GPT-5.6 versus Fable 5 and ChatGPT versus Claude—the selection logic is the same, only the model names change. There is a separate review of using Fable in Claude Code, Three days with Fable 5, and a guide to launching Codex, OpenAI Codex: what it is and how to launch it.

Get to know Ailoot.