This chart compares the four frontier coding models we test, Claude Opus 5.5, Claude Fable 5.1, GPT-6 Astra and GPT-6.1 Sol, on the numbers that matter when you pay for them. The first table is live: it aggregates every build in our catalog, so it updates whenever a new brief is added.
| Model | Briefs finished | Median wall time | Output tokens, all briefs | Average API-equivalent cost per build | Runtime errors | Interrupted attempts |
|---|---|---|---|---|---|---|
| Claude Opus 5.5 | 7 of 8 | 1h 14m | 1.8M (+1 not recorded) | $14.53 | 0 | 5 |
| Claude Fable 5.1 | 8 of 8 | 36m 57s | 1.3M | $13.85 | 0 | 5 |
| GPT-6 Astra | 8 of 8 | 11m 29s | 340k | $3.90 | 0 | 0 |
| GPT-6.1 Sol | 8 of 8 | 12m 05s | 394k | $0.80 | 0 | 0 |
What do the four models cost and how much can they read?
The vendors' own model pages, checked on October 21, 2026. Prices are per million tokens.
| Model | Released | Input | Output | Cached input | Context window | Max output |
|---|---|---|---|---|---|---|
| Claude Opus 5.5 | Sep 22, 2026 | $4 | $20 | $0.20 | 1M | 128K |
| Claude Fable 5.1 | Sep 1, 2026 | $10 | $50 | $0.25 | 1M | 128K |
| GPT-6 Astra | Sep 4, 2026 | $10 | $50 | $1.00 | 1.05M | 128K |
| GPT-6.1 Sol | Sep 29, 2026 | $2 | $10 | $0.10 | 1.05M | 128K |
Sources: Anthropic pricing, Anthropic model overview, OpenAI GPT-6 Astra model page and GPT-6.1 Sol model page. For prompts over 272,000 tokens, OpenAI doubles the input price and raises the output price by half.
What did each model actually cost per build?
Far more than the price sheet suggests for the Claude models, because they read and re-read their work many times. This is the API-equivalent cost of every build, live from our run logs.
| Brief | Claude Opus 5.5 | Claude Fable 5.1 | GPT-6 Astra | GPT-6.1 Sol |
|---|---|---|---|---|
| SaaS landing page | $7.15 | $10.88 | $2.88 | $0.69 |
| Analytics dashboard | $13.63 | $12.80 | $2.79 | $0.54 |
| Balls in a spinning hexagon | $8.73 | $8.75 | $2.12 | $0.65 |
| First-person shooter | not recorded | $20.48 | $5.64 | $1.03 |
| Mobile finance app | $21.06 | $11.46 | $2.73 | $0.83 |
| One-lane MOBA | in progress | $24.76 | $7.27 | $1.12 |
| Open brief | $16.65 | $8.85 | $4.17 | $0.71 |
| Kanban board | $19.98 | $12.83 | $3.60 | $0.83 |
Opus 5.5's shooter run hit our 150-minute cap before Claude Code could report a total, so its cost is not recorded. Notice that Opus, despite the lower token price, costs about the same per build as Fable: it works longer and reads more.
How should you read this chart?
As three trade-offs, not a ranking. Wall time includes each agent opening its own page in a browser and checking it, which the Claude runs do far more often. Output tokens measure how much each model wrote, including its reasoning. Cost tells you what a pay-as-you-go user would have paid. None of it tells you whether the result is good; for that, open the briefs and look.
Which model should you pick from the numbers alone?
GPT-6.1 Sol if budget decides: per build it was three to five times cheaper than Astra and ten to twenty-five times cheaper than the Claude models. Claude Opus 5.5 if you will pay for the most thorough first draft, since it outspends Fable only slightly while being Anthropic's recommended default. GPT-6 Astra is hard to justify on these numbers: it costs three to five and a half times as much as Sol per build, and on our briefs its output looked much like Sol's. Fable 5.1 is the premium pick for long, hard work, and the model reviews cover where that premium shows.
Numbers done. Now judge the work itself: four dashboards, names hidden.
Take the blind test
Frequently asked questions
Which AI model is cheapest for coding?
GPT-6.1 Sol, by a wide margin. At $2 per million input tokens and $10 per million output tokens it is the cheapest of the four on paper, and its builds cost between $0.54 and $1.12 each at list prices across all eight briefs, against about $7 to $25 for the two Claude models.
Which AI model has the largest context window?
They are effectively tied. GPT-6 Astra and GPT-6.1 Sol list 1,050,000 tokens and both Claude models list 1 million, each with up to 128,000 output tokens, so context size is not what separates them on projects like these.
What does API-equivalent cost mean in this chart?
Our runs used subscriptions, so nobody was billed per token. The figure is what each run would have cost at the vendor's published API prices, taken from Claude Code's own cost report for the Claude runs and calculated from fresh, cached and output tokens for the Codex runs.
Why is there no score column?
Because a single number hides the trade-offs the chart is meant to show, such as speed against polish. Every figure here is a measurement. To judge quality, watch the builds or take the blind test.