Public coding benchmarks leak into training data and rarely look like the code teams ship. So we built our own: 129 real changes from the Go backend that runs Requesty, with hidden tests the models never see.
Every model runs through the same coding agent, with the same tools and the same 80 turn budget, and every call goes through Requesty. The full results, with every metric we measure, are live on the coding benchmark page.
Who solves the most
claude-opus-5-5 leads at 92%, followed by claude-fable-5.1 (89%), gemini-4-argon, claude-sonnet-5-5 and gpt-6-astra (88%). gpt-6.1-terra (83%) and kimi-k3.1-pro (81%) are the strongest of this month's new releases.
Who is fastest
Score is half the story for coding agents. glm-5.3-flash returns a turn in 1.0s at the median and particle-1 in 2.0s, while gemini-4-argon writes the most output tokens per second (264).
Who needs the fewest tokens
claude-sonnet-5-5 processes 0.87M tokens per task, the fewest on the board, and finishes in 13 turns on average. Fewer turns and fewer tokens mean lower latency and lower bills.
Open weights
mistral-large-4 is the strongest open weight model at 78%, ranked #10, ahead of several closed models.
Try them
Every model on the board is available through one API on Requesty, including EU routing for teams that need it.
Frequently asked questions
- What is the Requesty coding benchmark?
- 129 private tasks taken from our own Go backend, from easy fixes to extreme multi file changes. Every model gets the same coding agent, the same tools and an 80 turn budget. A task counts as solved only when hidden tests pass and the existing tests still pass.
- Which model scores highest?
- claude-opus-5-5 solves 92% of tasks. The top six models are within six points of each other.
- Which open weight model is best?
- mistral-large-4 solves 78% of tasks, ranked #10 overall. deepseek-v4.1-flash follows at 77%, ranked #11.
- SEP '26
GLM-5.3 Flash vs DeepSeek V4 Flash 0731: which flash model fits your API workload?
GLM-5.3 Flash brings vision and stronger agentic scores; DeepSeek V4 Flash 0731 starts responding three times sooner. Which flash model fits your API workload?
- AUG '26
Claude Opus 5 vs GPT-5.6 Sol: API speed, coding and cost
Sol starts responding sooner. Opus streams faster and recorded fewer server errors. Compare coding results, context limits and costs after caching.
- JUL '26
Opus 5, Grok 4.6, GPT-6 rumors: shipping through the model release treadmill
Claude Opus 5 landed this week, Grok 4.6 is two weeks out, GPT-6 is rumored inside six, and DeepSeek and GLM keep shipping. If every launch costs you an engineering week, the problem is your architecture, not the release cadence.