Stop asking which model is best in general. Pick the model that fits the job: coding agents, repo refactors, frontend generation, code review, test generation, Chinese coding workflows, or budget-sensitive automation.
The right choice depends on workflow, public benchmark evidence, API pricing, context needs, retry risk, and output length. Use this page to shortlist models by scenario, then compare source-backed rows and estimate task cost before committing to a model.
Candidates for multi-step coding-agent loops need public coding evidence, visible retry caveats, and source-backed token pricing where available.
This is a shortlist label, not a universal winner. Estimate task cost before committing to a default model.
Compare candidatesRepo-level refactor and code review need editing evidence, context caution, and visible source freshness before cost comparison.
Advertised context is not the same as reliable long-context repo editing.
Compare candidatesFrontend generation needs coding evidence plus human inspection of UI output, follow-up rate, and implementation cost.
Preference or coding scores do not guarantee pixel-perfect UI. Treat the label as a workflow lens.
Compare candidatesLow-cost automation starts with token price, then checks benchmark coverage, retry rate, cache behavior, and cleanup cost.
Do not call a model cheapest overall unless task assumptions are visible.
Compare candidatesKimi, Qwen, DeepSeek-style rows can be useful when the page separates pricing, benchmark coverage, context, and unknown fields.
Partial evidence is acceptable. Fake certainty is not.
Compare candidatesUse the evidence table to see what supports each recommendation. The table should show benchmark/source, price availability, context status, data status, confidence, caveat, and calculator handoff. Recommendations can be reordered as data improves, but unsupported claims should not be shipped.
Strong SWE-bench evidence in captured source, but expensive output pricing; do not present as universal best model.
Pricing source: Anthropic Claude API pricing · checked 2026-06-02 · confidence high
Benchmark source: SWE-bench Leaderboards · checked 2026-06-02 · confidence medium
Good candidate for default coding workflow shortlist, but scenario label must cite benchmark and price fields rather than claim overall superiority.
Pricing source: Anthropic Claude API pricing · checked 2026-06-02 · confidence high
Benchmark source: SWE-bench Leaderboards · checked 2026-06-02 · confidence medium
Lower token price does not guarantee lower task cost if retry rate rises; calculator must expose retry assumptions.
Pricing source: Anthropic Claude API pricing · checked 2026-06-02 · confidence high
Benchmark source: SWE-bench Leaderboards · checked 2026-06-02 · confidence medium
OpenAI API pricing is source-backed for GPT-5.4 mini standard short-context pricing. Exact public coding benchmark for this alias is not verified, so benchmark evidence is marked not publicly benchmarked.
Pricing source: OpenAI API pricing · checked 2026-06-02 · confidence high
Benchmark source: SWE-bench Leaderboards exact alias recheck · checked 2026-06-05 · confidence low
This leaderboard does not combine SWE-bench, Aider, LiveCodeBench, arena scores, pricing tables, and usage signals into one fake universal score.
Pricing data prefers official provider pages. Aggregator or route-specific pricing must be labeled before it is used for a calculator prefill.
TTFT and tokens/sec only ship when a public source exists for the exact model or route. Otherwise speed is not disclosed.
A cheaper token price can become expensive if retries, output length, cache behavior, or failure cleanup change the workflow.
There is no single best LLM for every coding task. The best choice depends on your workflow, benchmark evidence, price, context, retry risk, and output length. Use scenario labels such as best for coding agents, best for refactor, or cheapest good-enough.
Start with models that have public coding evidence, reliable context behavior, visible caveats, and source-backed pricing. Then estimate the cost of your agent loop. Agent tasks can become expensive through retries, tool calls, long outputs, and failed trajectories.
The cheapest LLM for coding depends on task complexity and retry rate. A low price per million tokens can be attractive, but the real cost depends on how many prompts, outputs, retries, cache reads, and human corrections the workflow needs.
Compare them by workflow and source coverage. Separate API token pricing from subscription-tool pricing, keep benchmark sources separate, mark unknown fields, and avoid forcing all models into one absolute winner list.