The launch page calls GPT-6 Astra the strongest coding model. The invoice is a different story: list price is about 2.5× the last generation, and effort runs from low to max. What actually blocks a team is rarely “who is two points higher on a board.” It is one ticket: wall-clock time, tokens spent, the month-end bill, and whether the first diff is something you would merge.
The everyday shell is usually Codex. For how the products split, see Claude Code vs Codex vs Gemini 3.8 Flash; for how the money stacks, see how much an AI coding agent costs per month. Head-to-head with Fable on one project is in GPT-6 Astra vs Claude Fable 5.1: who really codes better. Below we only split Astra on four axes: speed, tokens, cost, completion.
What do public completion rates actually measure?
OpenAI’s September 2026 software-engineering table (knowledge cutoff 2026-04-30) looks roughly like this:
| Eval | GPT-6 Astra | GPT-5.6 Sol | Claude Fable 5.1 |
|---|---|---|---|
| Terminal-Bench 4.0 | 57.9% | 37.3% | 55.8% |
| DeepSWE v1.1 | 74.1% | 72.7% | 67.4% |
| Internal database-migration tasks | 63.9% | 42.7% | 57.8% |
| AA Coding Agent Index | 67.0 | 65.1 | — |
Artificial Analysis scores Astra 62 inside Codex, level with Fable 5.1 in Claude Code. The numbers can agree; the definitions do not. Official table, third-party harness, and “tests green plus a reviewer will merge” are three different completion rates.
How to read wall-clock speed and effort
Astra’s reasoning.effort is low / medium / high / xhigh / max. Higher effort means more loops, more browser checks, more “run it” instead of a blind apply-patch—wall clock and tokens rise together.
Two other wall-clock figures from the launch page:
- After the Codex harness update, Mind2Web finishes about 1.9× faster than GPT-5.6 Sol.
- OSWorld 2.0: Astra about 72.6% / 40 minutes, Sol about 65.7% / 75 minutes (roughly 47% less time).
- API Fast mode: about 2× standard speed, billed at about 2×.
On the same repo and the same “async CSV export” ticket, Astra more often greens the queue, retries, and CI first; empty states and loading copy usually need a second pass. A short wall clock is not a one-pass review.
Why token use and list price pull in opposite directions
List price moved from Sol’s $4 / $20 to $10 / $50 per million input / output—about 2.5×. Cache reads are $1, cache writes $12.50. You pay more per token; you often spend fewer tokens on the same job.
On Artificial Analysis in Codex at max, Astra uses about one-third of Sol’s tokens per task. Cost per task is about $7.09: roughly 40% cheaper than Fable 5.1 at the same score, only about 15% above Sol max, with a higher index. On the Intelligence Index, Astra max is about 27k output tokens; Fable often sits near 78k.
One reading rule: do not compare unit price alone. Short, execution-heavy jobs can eat the 2.5× sticker. Long sessions that reread the same large repo—and then trip the long-context surcharge—flip the bill.
What else is hiding in a single-task bill
| Line | Public list scale | When it doubles |
|---|---|---|
| Input / output | $10 / $50 | Fast mode ≈ ×2 on the whole request |
| Cache read / write | $1 / $12.50 | Input over 272K → long-context rates on the whole request (input and cache ≈ ×2, output ≈ ×1.5) |
| Seat | Inside the ChatGPT plan | Buy credits when the window is full; API has no cap unless you set one |
Plus / Pro usually includes Astra quota. That is not unlimited max. A mid-size agent turn that stuffs the repo, logs, and failure traces into the window hits the 272K surcharge before it hits the $10 row. If the machine sleeps, Prompt Cache dies and “continue” is billed as a full input—the most wasteful warm-up. How four ledgers stack for the month is still in the cost piece.
Where completion drops in a real repo
Write “real project completion” as three columns instead of one board:
- Eval complete: Terminal-Bench jumps hard versus Sol; DeepSWE is only +1.4—short patches were already close.
- Demo complete: queue, auth, retries—Astra often raises its hand earlier.
- Merge complete: empty states, disabled states, cross-module risk notes still often need a human line or another round.
The usual drop points: review stares at the UI; round two touches shared queries or cache keys without “read-only first, then a minimal diff”; people truncate context to save tokens and drop a constraint. Third-party write-ups also show a low-effort mid-size migration finishing in about half an hour for about eleven dollars. That is “it ran,” not “merge without reading.”
How to set effort and caps day to day
- Default medium or high. Keep max for cross-module refactors; use low to scout and write repro scripts.
- Do not make Fast the overnight default. Fine for a demo rush; a long refactor multiplies the unit price again.
- Stay under 272K when you can.
/clearbetween tickets is cheaper than/compacton a wall of failed logs. - Keep the machine awake. One hour with the lid shut voids the cache; the next turn is a full input.
- Score three columns. Green tests alone let the leaderboard inflate completion.
# three columns on the same ticket
wall-clock minutes | input / output / cache tokens | safe to send to review?
If you are timing speed and tokens, do not let the machine drop first
Astra’s wall clock and cache bill break when the lid closes: session dies, sandbox clears, Prompt Cache expires, the next turn is a full input, and the 272K surcharge can fire for no good reason. A Mac mini idles around 4W—enough to let Codex finish at your chosen effort instead of rebooting with the laptop.
Apple Silicon unified memory is a better fit for indexing a mid-size repo and running tests in parallel. macOS ships Unix, Homebrew, Docker, and SSH, so the terminal agent skips a WSL layer. Versus a same-price Windows box: fewer crashes, safer unattended runs, and Gatekeeper plus SIP are a better default for a long-lived agent.
If you are going to log wall clock and tokens for real, keep both sides the same SKU and awake—see the plans, and leave the difference in the model and the effort slider, not in sleep and re-warm.