Task guide · coding · 2026
Best AI for coding in 2026
For coding in 2026, pick the model and the editor separately. In this catalog GPT-6 Astra is 74.1% Pass@1 at $4.43/task on DeepSWE 1.1 (xhigh effort, best published row); Gemini 3.8 Flash is 73.8% Pass@1 at $2.36/task on DeepSWE 1.1 (high effort, best published row); Claude Opus 5.5 is not yet on Datacurve's independent DeepSWE 1.1 board; GPT-6 Luna is not yet on Datacurve's independent DeepSWE 1.1 board. Gemini 3.1 Pro is 11.7% Pass@1 at $2.14/task on DeepSWE 1.1 (high effort, best published row) — this is one benchmark result, not a general-quality ranking; validate the model on your agentic workload before choosing it. Prices below show standard short-context API rates; see each profile for context tiers, schedules and promotions. Dates reflect profile metadata.
Latest recorded profile verification:Source dates and limitations are shown on each model profile.
| Model | Tier | API price / 1M | DeepSWE | Input context / max output |
|---|---|---|---|---|
 Claude Opus 5.5 | Frontier | $4 in / $20 out Standard $4/$20 per MTok with no long-context surcharge. Cache reads $0.20; 5-minute writes $5 and 1-hour writes $8. Batch $2/$10. Fast research preview $8/$40 on the first-party Claude API only, not Batch. US-only inference adds 10% to all token categories where supported. Batch supports 300K output with output-300k-2026-03-24; standard output stays 128K. Knowledge cutoff Jun 2026; retirement not sooner than September 22, 2027. | Not on board yet | 1M context · 128K max out |
| Claude Fable 5.1 | Frontier | $10 in / $50 out Standard $10/$50 per MTok with no long-context surcharge. Cache reads $0.25, versus Fable 5's $1; 5-minute writes $12.50 and 1-hour writes $20. Batch $5/$25. US-only inference adds 10% where supported. Adaptive thinking is always on, default effort high; knowledge cutoff Jun 2026. Retirement not sooner than September 1, 2027. | Not on board yet | 1M context · 128K max out |
 GPT-6 Astra | Frontier | $10 in / $50 out Standard $10/$50 per MTok (≤272K input); above 272K the full request bills $20/$75. Cached input $1/$2 and cache writes $12.50/$25 for short/long context. Batch/Flex are half Standard; Fast is 2x. Ultrafast is 6x: $60/$300 short, $120/$450 long, through Responses with global/US processing; no EU or other regional inference residency. Regional/FedRAMP processing adds 10% where eligible. Knowledge cutoff Apr 30 2026. Enterprise model access is disabled by default. | 74.1% · $4.43/task · xhigh effort | 1.05M context · 128K max out |
| GPT-6 Sol | Balanced | $2 in / $10 out Standard $2/$10 per MTok; cached input $0.20 and cache writes $2.50. Above 272K the full request bills $4/$15, cached $0.40 and cache writes $5. Batch/Flex are half Standard; Fast is 2x. Regional/FedRAMP processing adds 10% where eligible. Knowledge cutoff Apr 20 2026; launched September 22 in the API, ChatGPT Work and Codex. | Not on board yet | 1.05M context · 128K max out |
| GPT-6 Luna | Budget | $0.10 in / $0.50 out Standard $0.10/$0.50 per MTok, compared with GPT-5.6 Luna’s $0.20/$1.20. Cached input $0.01 and cache writes $0.125. Above 272K the full request bills $0.20/$0.75, cached $0.02 and cache writes $0.25. Batch/Flex are half Standard; Fast is 2x. Regional/FedRAMP processing adds 10% where eligible. Knowledge cutoff May 18 2026. | Not on board yet | 1.05M context · 128K max out |
 Gemini 3.8 Flash | Balanced | $0.75 in / $3.75 out Intro Standard $0.75/$3.75 per MTok through December 31, 2026; $1.50/$7.50 from January 1, 2027. Output includes thinking tokens. Cached input $0.075, rising to $0.15; storage $0.50/MTok-hour, rising to $1. Batch/Flex input/output are $0.375/$1.875, rising to $0.75/$3.75. Priority $1.35/$6.75, rising to $2.70/$13.50. Grounding tools have separate charges and quotas; free-tier limits differ from paid. | 73.8% · $2.36/task · high effort | 1.05M context · 66K max out |
 Grok 4.7 | Frontier | $2 in / $6 out Standard $2/$6 per MTok below 200K prompt tokens; at 200K or more the full request bills $4/$12. Cached input $0.50 short or $1 long. US regional inference adds 10%. No Batch API. The separate Cursor/Grok Build-only Fast option costs 2x short rates or 1.5x long rates, billed through those plans; unavailable on the public xAI API. No separate text-output cap is published. Knowledge cutoff May 2026. | Not on board yet | 500K context · Not verified max out |
 DeepSeek V4.1 Flash | Balanced | $0.30 in / $1.20 out Official peak rates $0.30/$1.20 per MTok; off-peak $0.15/$0.60 (50% of peak). Cache hit $0.006 peak / $0.003 off-peak. Peak hours are weekdays 01:00–04:00 and 06:00–10:00 UTC, excluding Chinese public holidays; weekends, holidays and all other hours are off-peak. Open weights (MIT). The legacy deepseek-v4-flash and deepseek-v4-flash-vision-exp ids temporarily route here. | Not on board yet | 1M context · 384K max out |
Displayed rates are standard short-context USD per million tokens, excluding batch and cache. Long-context tiers, schedules and dated promotions can change the billed rate; see each profile before budgeting. On DeepSeek’s first-party API, V4-Flash is retired; its legacy API ID temporarily routes to V4.1-Flash at V4.1-Flash rates. Confirm DeepSeek’s live rate table before budgeting.
When the cheaper one wins
GPT-6 Luna is cheaper on output at $0.50 per million tokens against $20 for Claude Opus 5.5 — about 40×. Use the cheaper tier for classification, extraction, summarisation, and any task where the expensive model’s extra score does not change the accepted output. The expensive one only pays if your hardest task actually fails on the cheap tier. This compares standard short-context token rates, not total cost per accepted task. Include input, cache, context tiers, schedules, retries and review in your pilot.
Run the model pickerThe model is not the editor
Cursor, Claude Code, Copilot, and similar products wrap a model. Compare them on harnesses. A CursorBench score does not pick Claude Opus 5.5 versus GPT-6 Luna.
AI Model Picker is a free, no-signup tool that shortlists frontier, balanced, or budget models from catalogued API prices and context windows. If you already know you need an editor or terminal agent, use the coding assistant finder.
How to read DeepSWE here
Pass@1 alone can hide cost differences. Claude Opus 5.5 is not yet on Datacurve's independent DeepSWE 1.1 board. GPT-6 Luna is not yet on Datacurve's independent DeepSWE 1.1 board. The cheaper row wins when the extra points do not change the accepted patch.
Public DeepSWE 1.1 figures are mirrored from Datacurve. Vendor-only scores stay off this table.
Pair compare with the cheaper-wins rule: Claude Opus 5.5 vs GPT-6.1 Sol. Three-way brands: ChatGPT vs Claude vs Gemini 2026.
Common questions
Questions about AI for coding
Answered from the verified figures on this page rather than general guidance.