thirty-two open coding models, scored on eight independent benchmarks and blended with what each one actually costs per request on opencode go. no estimated cells: every number is quoted from a named source — datacurve, vals ai, artificial analysis, epoch ai or the vendor's own report.

8 benchmarks · 32 models ranked · 597 sourced values · también en español

• evidence report markdown · 96 kb
• full spreadsheet xlsx · 67 kb

the ranking

all 32 models currently on opencode go, best to worst on the benchmark-plus-cost blend. watch the cover column: a high rank built on four benchmarks is weaker than one built on seven.

how to read this

1. rank averages each model's rank across the eight benchmarks, blended 50/50 with its rank by cost per request. the weight is adjustable in the spreadsheet.
2. light blue marks the best value in each column across the 32 models.
3. — means no published value could be found anywhere as of 10 oct 2026. nothing is estimated.
4. cost is the estimated price of a typical go request (the token mix published in the go docs), deepseek at off-peak rates, tiered models at their base tier.

the benchmarks, briefly

eight benchmarks decide the ranking; the rest is context. each entry names who runs the eval, because the harness matters as much as the score.

the eight scored benchmarks

deepswe v1.1 datacurve

long-horizon software-engineering benchmark: 113 original tasks across 91 open-source repos and 5 languages, graded by program verifiers. tasks are written from scratch rather than scraped from real pull requests, which keeps contamination down; the public leaderboard runs every entry on one harness (mini-swe-agent), so the numbers are comparable.

terminal-bench 2.1 vals ai · terminus 2

89 sandboxed terminal tasks — package management, builds, git, servers, system administration. the 2.1 revision repaired 28 tasks to close reward-hacking gaps. we read vals ai's official terminus 2 runs as the comparable line; vendor numbers on the same benchmark are kept in the workbook but sit lower and must not be compared across that boundary.

scicode artificial analysis

research-grade scientific computing — physics, math, biology, chemistry problems broken into sub-steps requiring real domain background. artificial analysis publishes the current values; the underlying dataset is under review after an audit, so all scicode figures may be revised.

aa-lcr artificial analysis

long-context reasoning: 100 questions over 10,000–100,000-token documents, llm-judged. tests whether a model genuinely uses long context or just skims it — the failure mode that separates real context windows from spec-sheet context windows.

gpqa diamond artificial analysis · epoch ai

the diamond subset of google-proof q&a: expert-written graduate-level questions in physics, chemistry and biology, built so you can't guess your way through. values come from artificial analysis where available, epoch ai for the handful of models aa doesn't cover.

swe-bench pro mostly vendor-reported

scale ai's harder successor to swe-bench: longer, more realistic engineering tasks with multi-file changes. no standardized public board was reachable for most of these models, so this column is largely vendors reporting their own numbers — read it directionally, not as a leaderboard.

terminal-bench 4.0 vals ai

the current generation of terminal-bench — substantially harder than 2.1 (the best go score is 41.9) and highly sensitive to which agent ran it. independent mini-swe-agent numbers are preferred; vendor figures appear only where nothing independent exists.

mmlu-pro vals ai

the 14-choice, harder sibling of mmlu across 57 subjects — broad professional and stem knowledge with tougher distractors. vals ai has retired it for new model releases, so newer models simply have no value; that's a coverage gap, not a zero.

reference data — shown, not scored

aa coding index artificial analysis

the coding agent index: aa's composite of deepswe v1.1 + terminal-bench 4.0 + swe-atlas-qna, produced behind a named agent/harness (claude code, codex…). where the agents page is gated, values fall back to aa's per-model coding index — a different composite. the sources tab labels which product each figure is.

vals index vals ai

vals ai's versioned composite across a 15-sector suite: coding (incl. tb 4.0, vibe code bench, code migration), web, finance and more. the 2026-09-25 v2.1 revision reshuffled the mix, so only current-page values appear here — older-version readings are excluded rather than mixed in.

math per vendor

whichever math benchmark each vendor published: aime 2026, matharena apex, frontiermath tiers, hmmt, proofbench. the test differs per model, so the column is informational — the workbook's sources tab names the test behind each figure.

deepswe run metadata datacurve

the ± confidence interval, steps per task and output tokens per task from datacurve's board rows, where the model has one. context rather than a score: it shows whether a lead is real or inside the noise — gpt-6 astra's 74±3 against muse spark 1.3's unmeasured 75.4, for instance.

what stands out

the free model wins

step 5 preview free tops the ranking: three of the eight best-or-near-best scores (tb 2.1 85.0, scicode 58.9, aa-lcr 88.3) at zero cost. free previews are limited-time, so the practical pick is different: see muse spark 1.3 below.

a $0.10 model is the best value on the plan

muse spark 1.3 contributor is #2 overall and holds the deepswe crown (75.4) at roughly a quarter of a thousandth of a dollar per request. deepseek v4.1 flash (#3) pairs the lineup's second-best deepswe score, 74.2, with near-muse pricing. caveat: the contributor tier trains on your prompts and is region-limited.

terminal-bench 4.0 is a different world

the whole lineup struggles on the current terminal benchmark: the best go score is glm-5.3's 41.9. independent mini-swe-agent runs typically land 10+ points below vendor numbers, so treat every 4.0 figure as harness-bound.

raw scores don't win — cost does

kimi k3 leads aa-lcr (88.7) and scicode is nearly locked down by mimo-v2.6-pro (60.9), yet both sit mid-table: at $3/$15 and even $0.435 per million, their per-request cost drags the blend. pure-quality rankings would look very different.

composite indices, math & run details

informational columns from the spreadsheet — the composites and the math scores are not part of the ranking score. math uses a different benchmark per vendor (aime 2026, matharena apex, frontiermath, hmmt, proofbench…), so compare those with care.

not ranked

longcat 2.5 preview free and space bunny are on the plan but have no published scores on any of the eight benchmarks — space bunny's identity is still unconfirmed (a stealth listing, possibly minimax-family). until either publishes, they can't be placed.

no longer on the plan

kept for reference: seven models that appeared in the previous version of the sheet but are not in the current go lineup. three of them still exist pay-per-use on opencode zen.

changes

• 14 october 2026 two sentences under what stands out disagreed with the ranking, and were made to agree. no number in a table was touched. muse spark 1.3 contributor costs about a quarter of a thousandth of a cent dollar a request, and deepseek v4.1 flash pairs the best terminal-bench 2.1 score in the lineup the lineup's second-best deepswe score with near-muse pricing.

notes & caveats

1. swe-bench pro is nearly all vendor-reported. no standardized public board was reachable for most of these models; the few independent values (claude haiku 5.5 64.8, kimi k3 63.3) are aggregations. read the column directionally, not as a leaderboard.
2. the aa coding index mixes two products. where aa's coding agent index (deepswe v1.1 + terminal-bench 4.0 + swe-atlas-qna, harness-tagged) was readable, it is used; otherwise aa's per-model coding index — a different composite — stands in. the sources tab carries the exact product per value.
3. the vals index is versioned. current-page values are used (v2.1 where it exists); pre-v2.1 numbers (e.g. muse spark 1.3's 59.47-era readings) systematically differ and several model-page "accuracy headlines" were verified to not be the index at all and were excluded.
4. terminal-bench 4.0 is extremely harness-sensitive — vals' mini-swe-agent runs are preferred, vendor numbers kept only where nothing independent exists. minimax m3 illustrates it: 1.01 independent vs 66.0 vendor, on a different version of the benchmark.
5. deepswe ±ci / steps / output tokens come from datacurve's leaderboard rows only; they are missing wherever the model's row isn't one of the 21 rendered configs. use them to judge whether a lead is real (e.g. gpt-6 astra's 74±3 vs muse spark 1.3's unmeasured 75.4).

opencode.ai/go · generated 10 october 2026 from the updated benchmark workbook