T3 Code: real tasks, tokens, and cost
Landed work from Theo's fleet, July–October 2026. Each observation is a matched task, not a synthetic benchmark.
These models did different work. The lines compare the distribution of real task costs and sizes, not model quality or speed on the same task. A model label means it accounts for at least 90% of that task's estimated cost. Mixed work is available separately.
Cost per task
Landed change size
Model summaries
Medians describe a typical retained task. Each recorded model uses its own price; tasks were not re-run on other models.
| Model | Tasks | Median cost | Median lines | Median ¢ / line | Median tokens / line | Cache share |
|---|---|---|---|---|---|---|
| GPT-6 Astra | 8 | $2.76 | 168 | 2.92¢ | 19,047 | 95.7% |
| GPT-5.6 Sol | 34 | $10.47 | 180 | 4.41¢ | 77,753 | 96.8% |
| GPT-6.1 Sol | 3 | $0.58 | 37 | 1.56¢ | 76,662 | 96.4% |
| Claude Fable 5 | 81 | $18.34 | 139 | 14.56¢ | 90,772 | 97.3% |
| Claude Fable 5.1 | 36 | $26.90 | 354 | 8.36¢ | 64,090 | 94.9% |
| Claude Opus 5.5 | 22 | $6.35 | 134 | 4.48¢ | 98,595 | 98.3% |
Cost versus change size
Select a task to see its measured usage and model cost split.
Task explorer
Select a title to inspect its model mix, token usage, and landed diff. PR links open the evidence on GitHub.
| Task / PR | Model | Lines | Tokens | API cost | ¢ / line |
|---|---|---|---|---|---|
| feat: native subagent & workflow observability#5219 | Claude Fable 5 | 6,942 | 947.47M | $1,226.53 | 17.67¢ |
| feat(mobile): one sheet for model and thread settings#5625 | Claude Fable 5 | 1,548 | 392.75M | $487.10 | 31.47¢ |
| feat: paginate thread loading with user-anchored turn windows#5493 | Claude Fable 5 | 2,123 | 267.41M | $406.74 | 19.16¢ |
| feat(mobile): pick, share, and receive files in threads#8237 | Claude Fable 5 | 7,756 | 226.98M | $353.54 | 4.56¢ |
| feat(sidebar-v2): thread snoozing#4311 | Claude Fable 5 | 2,624 | 204.44M | $312.21 | 11.90¢ |
| fix(swift-ios): apply audit findings across transport, home, chat, and settings#10843 | Claude Fable 5.1 | 3,229 | 142.99M | $262.52 | 8.13¢ |
| feat(web): per-device provider settings#4479 | Mixed models | 2,649 | 187.96M | $257.84 | 9.73¢ |
| feat(web): queue messages sent client-side while the agent is working#11673 | Claude Fable 5.1 | 869 | 227.07M | $252.98 | 29.11¢ |
| feat(desktop): update the desktop app on remote Macs from the Update button#6554 | Mixed models | 3,940 | 154.20M | $244.03 | 6.19¢ |
| fix(antigravity): refresh the model manifest so older Gemini models fold as legacy#9397 | Claude Fable 5.1 | 84 | 225.81M | $231.96 | 276.14¢ |
| fix(mobile): stop long iOS threads from jumping while scrolling up#4867 | Claude Fable 5 | 284 | 165.98M | $223.95 | 78.85¢ |
| feat(web): drag pinned threads into your own order#5581 | Claude Fable 5 | 1,499 | 142.32M | $210.17 | 14.02¢ |
How to read this
Measured work
- 682 older merged PRs were scanned. 178 had usable, focused task records. They are combined with the earlier 33-PR sample, for 211 unique PRs across 206 tasks.
- Branches and saved PR links join T3 Code threads to merged GitHub diffs. Shared tasks are grouped. Usage runs from task start through the final merge.
- Usage comes from native Claude and Codex records plus session- or response-matched provider captures. Repeated response records are counted once. Cache reads, cache writes, and output remain separate.
- Code + config excludes tests, fixtures, generated files, and prose, using T3 Code's own path classifier. A truncated file list is excluded from that view.
- The test purge, reverts, broad maintenance work, and records with unresolved usage or prices are excluded. Empty or missing usage is never treated as zero.
Limits that matter
- This is an observed sample, not an exhaustive accounting of every attempt. It includes matched takeovers and reviews, but cannot prove all prior work was recovered.
- Fable 5 mainly appears in July–August, Fable 5.1 in September, and Opus 5.5 in October. Task mix, instructions, software, and workflow changed over time.
- GPT-6 Astra has only eight retained tasks and GPT-6.1 Sol has only three predominantly single-model tasks. Do not infer a model ranking from those samples.
- The price calculation uses today's standard token rates for consistent comparison. It is not a historical invoice or a subscription bill. Fast/priority premiums, tool charges, and external review-bot costs are not included.
- “Tokens” includes repeated cached context. “Fresh tokens” means uncached input + cache writes + output; it still includes reasoning and tool context, not just code generated.
Price assumptions · USD per million tokens
| Model | Input | Cache read | Cache write | Output |
|---|---|---|---|---|
| GPT-5.6 Sol | $4 | $0.40 | $5 / $10* | $20 |
| GPT-6 Astra | $10 | $1 | $12.50 / $25* | $50 |
| GPT-6.1 Sol | $2 | $0.10 | $2.50 / $5* | $10 |
| Claude Fable 5 | $10 | $1 | $12.50 / $20 | $50 |
| Claude Fable 5.1 | $10 | $0.25 | $12.50 / $20 | $50 |
| Claude Opus 5.5 | $4 | $0.20 | $5 / $8 | $20 |
Claude cache writes: 5-minute / 1-hour. *OpenAI: 30-minute / 12-hour; no separately billed writes were observed for these OpenAI task records. Other recorded review models use their own rates. Sources: Anthropic pricing · OpenAI pricing. Rates checked October 20, 2026. Dates in the task data are UTC.
Chart colors, direct labels, and generous line hit areas take inspiration from Slopalytics. Analysis by GPT-6.1 Sol in T3 Code.