The top two are 0.4 points apart. This is no longer about who is smarter.
On August 21, an independent Terminal-Bench 2.1 score dropped. In the coding-agent track, the gap between first and second is now negligible: GPT-5.6 Sol scored 89.5%, Claude Opus 5 scored 89.1%.
The "best model" you pay a premium for may differ from the runner-up by about the time it takes to drink a coffee.
The real numbers
- GPT-5.6 Sol (xhigh): 89.5%, first
- Claude Opus 5 (max): 89.1%, second
- Grok 4.6 (high): 88.4%, third
- Claude Code + Fable 5 (vendor self-report): 83.8%
- Codex CLI + GPT-5.5: 83.1%
- Claude Code + Opus 4.8 (vendor): 78.9%
One trap to note: the top three are third-party Artificial Analysis runs on the same Terminus 2 harness across 89 real terminal tasks, so they compare. The lower ones are vendor-submitted and not directly comparable. Do not take a vendor poster score as gospel.
On the open side, star counts are the real battlefield
For coding agents, open-source activity tells the story. opencode pulled 199,640 stars, beating Claude Code's 142,130, Codex's 109,229, and Gemini CLI's 106,596. It is MIT-licensed, model-agnostic, free.
The price ladder is clear too: Cursor Pro at $20 a month, GitHub Copilot Pro at $10 gets you in. What is expensive is no longer "can it code" but "how well, and at what cost."
The real divide is cost and workflow
GPT-5.6 Sol lists at $5 per million input tokens, $30 out. Opus 5 is $5 in, $25 out. Half a percentage point separates them on the board, a wide gap on the bill.
Anthropic positions Opus 5 as the practical default, reserving Fable 5 for cases where the score gap justifies double the token price. OpenAI runs a three-tier family, Sol, Terra, Luna, across the board.
Plainly, the board has reached the point where it measures "what you will pay for half a point," not "who is absolutely best." The same model scores differently inside different agents, so what gets rated is the agent-plus-model pair, not the model alone. Swap your frontend tool and the score may move.
One overlooked fact
These scores test "drive a terminal and finish a dev task end to end": edit files, run commands, fix failures. They do not measure "how pleasant is writing business code." Day to day, when you build features and fix bugs, the feel often disagrees with the board. The board shows who has the higher ceiling, not who fits your mouse better.
The photo-finish is actually driving prices down
With the top two this tight, the direct consequence is a price war. Grok 4.6 launched at $2/$6 per million tokens precisely to undercut GPT and Claude by a wide margin. As model capability converges, vendors can only compete on price-performance, and that is good for us. A year ago, coding at this level was billed per token on GPT-4-class pricing. Now the same level costs several times less.
The practical upshot is that the leaderboard matters less for buying than it did a year ago. The top three sit within a point and a half, and the gap shrinks every quarter. What actually shows up in your bill and your daily flow is the tool wrapped around the model, the context window, the integrations, not the model name on the tin.
One more note for teams: because scores are this close, a bad day of prompting or a mismatched agent framework can swing your result more than the half-point between first and second. Benchmark the pair you will actually use, not the headline model.
Why this matters to you
When you pick a coding tool, stop worshipping the number one slot. A 0.4-point gap is invisible in daily use. What actually counts: does it fit your workflow, is the context window long enough, can you afford the bill. Small teams start at Copilot Pro for $10; heavy users move up to Opus 5 or GPT-5.6 Sol. Open-source folks take opencode and choose their own model. When scores are this close, price and feel are what you should agonize over.