GPT-6 Sol and Luna vs Claude Opus 5.5: Which Costs Less to Finish Real Tasks?

What a passing code patch cost in the same test

FrontierCode Main contains 100 tasks from open-source repositories. A run passes when its patch clears all maintainer-defined blocking criteria; the separate score also accounts for non-blocking quality criteria. Cognition ran each model five times per task and effort level. Its leaderboard data reports pass rate and cost per rollout. The last column divides the reported average rollout cost by the pass rate, so it represents aggregate token spend per benchmark pass, including failed runs. It does not include a repair attempt or a human review bill.

At low effort, Opus produced more passing patches than Sol and cost slightly less per run, despite its higher list price per token. At medium effort, Sol's $0.7659 per run was below Opus's $0.8016, but Opus's higher pass rate reversed their cost per pass: about $1.34 against $1.51. Luna remained cheapest among these settings even at max effort, though its 47.8% pass rate and 42.4% rubric score were below Opus at medium. Moving Luna from low to max lifted its pass rate by 18.6 points and roughly tripled spend per pass.

These figures compare maintainer-designed coding tasks under Cognition's grading rules. The GPT models ran in Codex CLI and Opus ran in Claude Code, so differences can come from the agent harness as well as the model. Cognition's pass means that no blocking criterion failed; it is not a recorded production merge. A different repository, acceptance bar, prompt, tool setup, or human repair process can change the ordering.

For cross-app workflows, Sol and Opus reverse order

Zapier's AutomationBench 1.0.6 leaderboard uses held-out business workflow tasks. The agent searches and changes records across simulated CRM, inbox, and spreadsheet tools; a task passes only when every final-state assertion holds. The table divides Zapier's rounded cost per attempted task by its strict completion rate. These are approximate model charges per benchmark completion, not measured retry sequences.

Luna high had the lowest calculated spend per completion, but only 14.5% of its runs finished the whole workflow; max raised that to 20.7%. Sol xhigh and Opus high finished almost the same share, while Sol's reported run cost was $0.27 against $0.71. Opus max reached 42.47%, a level neither Sol nor Luna reached in this leaderboard, at a higher bill. Zapier's Opus rows use Anthropic's default fallback routing for refused steps, so those rows describe that setup. The per-run costs are rounded to cents; the calculated spend per completion is correspondingly approximate, especially for Luna.

One illustrative Zapier task asks the agent to close the right Meridian deal and send a notice to the right teams. Three near-identical CRM names, stale exchange-rate and account-tier rows, and a routing policy hidden in email all matter. Updating the wrong deal or notifying the wrong team fails the strict check even if the agent says it finished. That is why a cheap API call and a plausible reply do not establish a cheap completed workflow.

Why the API price card gives a different impression

These are the providers' OpenAI and Anthropic Standard API rates in US dollars per million tokens, checked September 24, 2026. The OpenAI rates apply when the prompt has at most 272,000 input tokens. They price tokens, not a completed task.

Sol's new-input and output rates are 20 times Luna's; Opus's are twice Sol's. Yet FrontierCode's low-effort runs cost $0.4036 on Opus and $0.4324 on Sol. Their token usage differed, so the per-token ratio did not predict the run bill. Subscription limits, Codex credits, cloud marketplaces, Batch, Fast mode, and paid tools have separate billing rules; the Codex cost guide covers credits.

Watch the 272,000-token price step

The illustration holds token counts equal, assumes all input is new, and excludes tools and retries. A 20,000-input, 2,000-output call costs $0.003 on Luna, $0.060 on Sol, or $0.120 on Opus. At 300,000 input and 10,000 output tokens, the charges become $0.0675, $1.350, and $1.400. These are rate-card calculations, not observed task bills.

Once a Sol or Luna prompt exceeds 272,000 input tokens, OpenAI charges twice its input rate and 1.5 times its output rate for the whole request. Sol's long-call calculation is (0.3 × $4) + (0.01 × $15) = $1.35. Opus 5.5 retains standard pricing across its 1M-token context window, giving (0.3 × $4) + (0.01 × $20) = $1.40. The short-call price ratio does not survive this threshold; cache use and different output lengths can move the actual bills again.

Apply the result to your own work

For code changes, define the review blockers before running the models: missing behavior, broken tests, an unrelated file change, or a violated project convention. For a cross-app workflow, check the final records and required notifications rather than the agent's reply. Run the same held-out jobs through the settings you might buy, then record the agent harness, billed charges, and pass or fail for every attempt. Compute total run charges ÷ accepted jobs. If you repair a failure, include the repair run and reviewer labor before counting that job as accepted.

All-in dollars per accepted task =
  (model and tool charges
   + validation charges
   + reviewer and repair minutes × hourly labor rate / 60)
  / accepted tasks

For document extraction, research, or another job outside these two benchmarks, use that job's own acceptance checks. Neither leaderboard supplies its pass rate. If none of your tested runs passes, report the spend and zero accepted tasks; there is no finite cost per accepted task to calculate.