Coding Model Rankings 2026

Compiled 2026-09-24. The main table copies the official Terminal-Bench 4.0 board; the board record was updated on 21 September 2026. Each row is one submission: model, reasoning effort and agent are scored together. GPT-6 Astra (max) with Codex resolves 58.2% of tasks; the 95% half-width is 2.8 points. Fable 5.1 (max) with Claude Code resolves 57.9%, half-width 3.8 points, and the two intervals overlap. Coding products: AI coding assistants. Crowd votes: LLM capability. List prices: LLM API pricing.

Terminal-Bench 4.0 official board

These are all 27 rows on that board. Tied ranks stay as printed. Cost is the bill for that evaluation run, not a per-million-token list price. GLM-5.3 (max) with Claude Code is 16th at 41.8%. Grok 4.7 (xhigh) with Grok Build was submitted on 21 September 2026 at 37.6%.

SWE-bench Verified, one agent: mini-SWE-agent 2.0.0

The public SWE-bench board stacks different agents. The table below keeps bash-only submissions on mini-SWE-agent 2.0.0 and sorts them by resolve rate. Claude 4.5 Opus (high) is at 76.80%, average cost $0.75, submitted 17 February 2026. MiniMax M2.5 (high) and Gemini 3 Flash (high) are both at 75.80%. Kimi K2.5 (high) is at 70.80%, average $0.15. DeepSeek V3.2 (high) is at 70.00%.

LiveCodeBench default window (through 1 May 2025)

The default board lists 454 problems dated 1 August 2024 through 1 May 2025. o4-mini (high) has Pass@1 80.2, Easy 99.1, Hard 63.5. DeepSeek-R1-0528 is 5th at 73.1. That generation stops at the window’s end date. Astra, Fable 5.1 and Opus 5, released later, are absent.

How this list is ranked

Compiled 2026-09-24. The main table is the public board titled Terminal-Bench 4.0 on tbench.ai. The board record was updated on 21 September 2026. All 27 rows are included. Resolve rate, 95% half-width, cost, tokens and submission date are copied from that board’s JSON. Cost is total_cost_usd rounded to the dollar. The second table keeps only SWE-bench Verified rows that use bash-only and mini-SWE-agent 2.0.0, sorted by resolve rate. Those ranks are internal to the 12 rows. The third table is the LiveCodeBench default window, problems dated 1 August 2024 through 1 May 2025 (454 problems). That generation stops in May 2025 and is not folded into the main order.