Claude Fable 5.1 and GPT-6 Astra are both closed, cloud-only frontier models. That is where the similarity ends. Astra's announcement leads with computer use, abstract reasoning, and office automation. Anthropic's Fable 5.1 announcement leads with long-running research, coding verification, and a cheaper way to reuse context.
This is not a controlled head-to-head test. The figures below come from the two labs' announcements, with different harnesses, dates, and evaluation choices. Use them to understand the models' intended strengths, not as a single league table.
The quick verdict
- Choose Fable 5.1 for: long coding sessions, research orchestration, root-cause analysis, writing quality, and workflows that repeatedly reuse a large context.
- Choose GPT-6 Astra for: computer control, browser and desktop automation, abstract puzzle solving, and office workflows where fast action matters.
- Neither is local: neither model can be downloaded or run on an AMD GPU.
- Price is deceptively similar: both list $10 per million input tokens and $50 per million output tokens, but Fable's much cheaper cache reads can make repeated-context jobs cheaper.
Published numbers, side by side
| Workload | Claude Fable 5.1 | GPT-6 Astra | What to notice |
|---|---|---|---|
| Agentic coding | 55.8% Terminal-Bench 4.0 73.4% CursorBench |
57.9% Terminal-Bench 4.0 | Close on one terminal benchmark; harness details matter. |
| Computer use | 77.9% OSWorld partial 41.7% strict |
72.6% OSWorld 2.0 | Not directly comparable releases or scoring rules. |
| Multidisciplinary reasoning | 65.0% Humanity's Last Exam with tools | 57.2% with tools in OpenAI's comparison | Fable looks stronger on this published slice. |
| Knowledge work | 1853 GDPval-AA v2 | Strong office and automation claims | Different evaluation styles, not a clean match. |
| Abstract reasoning | Not highlighted in the launch table | 99.9% ARC-AGI-3 | Astra's most dramatic published result. |
| Agent economics | Cache reads: $0.25/M tokens | Standard token pricing | Fable has the clearer repeated-context advantage. |
Fable is built for the long middle of a task
Many agents fail after the exciting first step. They lose why a fix failed, forget a constraint, or declare victory without checking the output. Fable 5.1's launch material repeatedly emphasizes the opposite behavior: durable notes, verification loops, and long unattended runs.
That makes Fable attractive for a codebase migration, a difficult incident investigation, or a research question that needs several experiments. The model's advantage is not that it always knows the answer immediately. It is that it is more willing to do the work required to find out.
Astra is built to operate the computer
OpenAI's Astra announcement is more concrete about the desktop: filling forms, updating a CRM, booking appointments, browsing listings, and routing a circuit board in KiCad. Its OSWorld and Mind2Web claims point at a model that can turn an instruction into a sequence of UI actions quickly.
If the job is “take this structured request and complete the workflow in five applications,” Astra's product direction is unusually clear. It is less about a brilliant paragraph and more about getting the clicks, fields, and checks done.
Research: insight versus execution
Fable 5.1's strongest stories are about insight: finding a rare crash, identifying a hidden gap in clinical research, optimizing GPU kernels, and running parallel experiments. Anthropic also describes scientific work such as a higher-resolution map of Venus and biology optimizations, though some of the most permissive life-science capabilities are limited to Mythos access programs.
Astra's strongest public story is execution: abstract reasoning, computer use, terminal work, and office automation. In practice, the best research assistant may be a combination: one model proposes and investigates, while another operates the software and turns the plan into a completed workflow.
Safety is part of the comparison
Both companies treat frontier capability and safeguards as linked. Anthropic separates Fable and Mythos access, allows defensive vulnerability discovery while restricting exploit development, and uses Enterprise Frontier Safeguards for customer-controlled data handling. OpenAI places Astra's cyber capabilities inside its Preparedness Framework and reports that the model found real vulnerabilities during testing.
Do not read a benchmark result as permission to automate dangerous work. A model that can find a vulnerability is useful only when the surrounding process controls access, validates output, and keeps a human responsible for consequential actions.
Which one should you use?
| Your task | First model to try | Why |
|---|---|---|
| Multi-hour coding or code review | Fable 5.1 | Long-horizon work, verification, and cache economics. |
| Browser or desktop automation | GPT-6 Astra | Its published product story is built around computer use. |
| Open-ended research | Fable 5.1 | Better fit for orchestrating sources, experiments, and notes. |
| Repetitive office workflows | GPT-6 Astra | Strong focus on forms, apps, and action sequences. |
| Private or offline work | Neither | Use an appropriate open-weight local model instead. |
My take
Astra feels like the sharper operator: hand it a computer and it wants to finish the workflow. Fable 5.1 feels like the stronger investigator: give it a messy, ambiguous problem and time, and it wants to understand the system before changing it.
The practical winner is therefore workload-dependent, and the best engineering choice may be to test both on your own tasks. The published numbers are useful clues, but your repository, browser, tools, latency budget, and failure tolerance are the real benchmark.