GPT-6 Astra and Claude Fable 5.1 scored within 0.3 points of each other on Terminal Bench 4.0, captured September 11. Astra ran it for $3,300. Fable ran the same benchmark for $6,200.
The scores are a statistical tie. The bills are not.
What Terminal Bench 4.0 Measures
This is not a writing test. Terminal Bench gives an AI agent a real computer task and steps back. No hints, no human correction in the loop. The agent finishes the job or it does not.
Astra ran the test via Codex. Fable ran it via Claude Code. Same class of agentic tasks, same measurement.
The results:
- GPT-6 Astra via Codex: 58.2%, margin of error ±2.8 points, cost $3,300
- Claude Fable 5.1 via Claude Code: 57.9%, margin of error ±3.8 points, cost $6,200
The 0.3-point gap sits inside the margin of error for both models. By any statistical measure, they performed identically.
The Column Nobody Screenshots
Most posts about this benchmark show the scores. Almost none show the cost column.
Same tasks. Same benchmark. Same result. Fable spent 88% more to get there.
That is not a flaw in Fable. It is a clue about what Fable is actually built for.
Anthropic describes Fable 5.1 as designed for long-running, asynchronous work. The jobs you start and walk away from. Rebuilding a codebase from a spec. Processing a long document overnight. Running a multi-step workflow that takes hours.
For those jobs, Fable's cache pricing matters. Cache reads on Fable 5.1 cost 75% less than on Fable 5. If your workflows re-read the same documents across sessions, that saving compounds fast.
Astra is faster and cheaper on short, contained tasks. One prompt, one output, done.
Neither model wins every job.
The Mistake That Costs Business Owners the Most
Picking one model and committing to it is not the practical choice. It is the expensive one.
Both models will reprice next quarter. Both will improve on some tasks and fall back on others. Building all your workflows on a single model means rewriting when it moves.
The smarter setup runs them in parallel. One model builds the output. The other reviews it from a clean context window. They never share chat history. Each catches what the other skips because they were trained on different data with different defaults.
This is the dual-review setup. It costs cents per run and survives any leaderboard change.
The Prompt: Copy This
Copy this.
You are a strict reviewer. I will give you an output produced by a different AI model. Your job is to find problems only. No suggestions, no praise.
Check for: factual errors, logic gaps, missing edge cases, and assumptions the model did not label as assumptions.
Return a numbered list of problems. If you find none, say: no problems found.
[PASTE THE OUTPUT HERE]
Open Codex after any Claude Code session where you are not certain about the output. Paste the output with this prompt. The reviewer reads it cold, with no context from the build session.
Switch the direction any time. Run Codex for the build, Claude for the review. The skill works in either order.
What You Get Back
- A numbered list of specific problems, not general feedback
- Logic gaps the builder model treated as obvious
- Factual errors that would have made the output wrong in edge cases
- Assumptions that were never labeled as assumptions
- A list you can act on immediately, not a list of compliments
How to Save It as a Skill
The full dual-review skill package is linked at the bottom of this page. It includes a pre-configured setup for Claude Code and Codex running in separate windows, with the review prompt pre-loaded.
The resource package linked here adds the connection setup for running the two models without context bleed. The prompt above runs standalone. The package adds the guardrails.
If you already have the dual-review skill from the earlier walkthrough, you have everything you need. This package is the updated version with Fable 5.1 and Codex 6 configurations.
Honest: When This Does Not Pay Off
For a one-sentence question or a quick lookup, running two models is overhead without a return. Skip the review for anything you would not show a client.
For client deliverables, production code, and any output that goes out under your business name: the extra step pays. You get a second opinion at API cost, not at the cost of a person's time.
One Action This Week
Pick one output you sent this week that you were not fully confident about. Paste it into the other model with the prompt above. Read the list.
If it finds a problem, you caught something before the next version ships. If it finds none, you spent five minutes and confirmed the output was clean.
That is the whole case for running both.