待翻译：GLM-5.3 vs. GPT-5.6 Sol on DeepSWE: Cost, Coding, and Routing

AI 服务暂时不可用，以下为来源正文，待恢复后补全翻译。

All blog posts Model Library Published 8/22/2026 GLM-5.3 vs. GPT-5.6 Sol on DeepSWE: Cost, Coding, and Routing Sol wins the first try, GLM-5.3 wins the rest at half the price, and the cascade beats both: 85.9% at \$6.61 a task. Authors Zain Hasan, Shobhit Dixit Table of contents 40+ Models Chosen for Production...40+ Models Chosen for Production...40+ Models Chosen for Production... Key Takeaways Don't pick one. Run GLM-5.3 first, escalate to GPT-5.6 Sol when the tests fail. That cascade solves 85.9% of DeepSWE tasks at \$6.61 each. Sol alone solves 72.7% at \$8.37. Thirteen points better, 21% cheaper. Sol wins the single shot, narrowly. 72.7% pass@1 against GLM-5.3's 69.0%, a 3.7 point gap that sits inside a couple of standard deviations. GLM-5.3 wins every retry after that. It ties Sol at pass@2 (81.1 vs. 81.0) and leads pass@4 (87.6% vs. 85.8%). The price gap is 2.1x. \$3.99 per rollout against \$8.37. Per \$100 spent, GLM-5.3 solves 17 tasks and Sol solves 9. Sol is faster and steadier: 19 minutes and 61 steps per rollout against GLM's 35 and 124, with 61 tasks solved four for four against 48. GLM-5.3's failures are cleaner. It breaks tests that already passed in 11% of its failures, against 20% for Sol. Gate Sol's diffs on regressions. The two diverge (0.43 per-task correlation) and cover 106 of 113 tasks between them, which is what makes the cascade work. We ran GLM-5.3 (max) against GPT-5.6 Sol (max) on all 113 DeepSWE tasks, four trials each, from the published per-trial records: 904 rollouts in total, 452 per side. Sol is the precision flagship. GLM-5.3 is the open-weight challenger that closed the gap. Every figure below comes from this run, so it can differ from other public GLM-5.3 vs. GPT-5.6 Sol scorecards. DeepSWE · Head to Head GLM 5.3 vs GPT 5.6 Sol at a glance Model Pass@1 Avg cost Solves / $100 Out tok Steps glm-5.3 [max] 69.0% ± 2.7% $3.99 17 80k 124 gpt-5.6-sol [max] 72.7% ± 2.2% $8.37 9 60k 61 GPT-5.6 Sol still holds the single-shot crown on DeepSWE, a benchmark that tests a model's software engineering ability across many task types and programming languages. GLM-5.3 arrives less than four points behind it at half the price and pulls ahead the moment you allow more than one attempt. This is the closest the open tier has come to the frontier, and the question worth answering is what Sol's remaining premium actually buys. The DeepSWE scoreboard: pass@1 and pass@k Single shot, Sol edges ahead: 72.7% pass@1 to GLM-5.3's 69.0% under DeepSWE's official scoring. Allow retries and the order flips. At two attempts GLM-5.3 (81.1%) already ties Sol (81.0%); at four, GLM-5.3's 87.6% pass@4 leads 85.8%. The open model has the wider reach, so in any best-of-k setting it is the more accurate choice, and its extra attempts cost half of Sol's. Cost comparison: GLM-5.3 vs. GPT-5.6 Sol pricing At \$3.99 a rollout, GLM-5.3 is 2.1x cheaper than Sol (\$8.37), which in value terms is 17 solves per \$100 against Sol's 9. Sol buys that premium back on latency: an average 19 minutes and 61 steps against GLM's 35 minutes and 124 steps, with 60k output tokens to GLM's 80k. The tradeoff is unusually clean. Sol is the faster, terser worker; GLM-5.3 is the lower-cost one that takes the long route. If a human is waiting, Sol earns its premium on latency alone. If a budget or a batch queue is waiting, GLM-5.3 is the better buy. Coverage vs. reliability: precision vs. reach Decompose pass@1 and pass@4 into coverage (tasks solved at least once across four tries) and reliability (tasks solved on all four), and the split is clean. Sol is the precision corner: 84.5% reliability and 61 tasks solved four for four, the marks of a model that lands what it touches. GLM-5.3 trades to the other axis: wider coverage at 87.6% against 85.8%, but lower reliability at 78.8% and fewer solid tasks, 48 against 61. That coverage edge is the same fact as its pass@4 lead. GLM-5.3 touches more of the benchmark than Sol and converts each touch a little less often per shot. Failure modes: how each model gets it wrong The failure profiles differ sharply, and the split favors the open model. Sol breaks the repository's existing test suite in 20% of its failures, the GPT-family regression signature. GLM-5.3 does so in 11%; when it misses, it usually misses forward, a near miss with the baseline intact, at a 61% near-miss rate against Sol's 54%. So the lower-cost model is also the safer one to accept without a heavy regression gate. Put a full regression run around Sol before you take its diff. GLM-5.3 needs that guardrail less. Where each wins, by task type The domain map splits evenly, four each. Sol takes data modeling and serialization (92%), build and ops tooling (73%), concurrency and durability (72%), and protocol conformance (59%), which is the exact-contract, systems-heavy work. GLM-5.3 takes query and config languages (88%, the highest single cell on the board), language and runtime internals (83%), stateful reactivity (73%), and program analysis, a 64 to 64 tie it holds on volume. That is the structured, interpreter-style work. GLM-5.3's one clear hole is protocol conformance at 44%, 15 points behind Sol. Sol has no single weak domain; it is simply strong across the board. Task types here were classified by an LLM from each benchmark prompt. GLM-5.3 vs. GPT-5.6 Sol by programming language GLM-5.3's standout is JavaScript at 90%, 15 points clear of Sol's 75 and the best JS number of any model on the board. It also takes Rust, 70 to 60. Sol answers with Python (74 to 66), Go (79 to 76), and TypeScript (66 to 61). The routing rule is simple: JavaScript and Rust to GLM-5.3, the rest to Sol, with GLM-5.3 a lower-cost and close second almost everywhere. How different are GLM-5.3 and GPT-5.6 Sol? Different enough to route on. Per-task correlation is 0.43, real disagreement for two models this close on aggregate. They both solve 90 tasks; GLM-5.3 alone gets 9, Sol alone gets 7, and 7 defeat both. Their union covers 106 of 113 tasks (93.8%), and the hard disagreements run one way: GLM-5.3 sweeps two tasks four for four that Sol never lands (koota-pair-relation-tracking, participle-grammar-conflict-analysis), while Sol sweeps none that GLM-5.3 zeros. The open model reaches places the flagship does not. Routing between them: the portfolio play That divergence plus the price makes the cascade the best row on the board. Run GLM-5.3 first and escalate to Sol only when your test suite rejects the answer: 85.9% solved at \$6.61 per task. That is thirteen points above Sol alone (72.7%) and still cheaper than one Sol rollout (\$8.37), because the open model clears most of the queue at half of Sol's price and the hard remainder gets a second independent attempt. The cascade also beats a perfect one-shot oracle router (83.8%), because two independent attempts beat one perfect pick. Expected accuracy is the same whichever model leads, but GLM-first is cheaper (\$6.61 against \$9.47), so lead with the lower-cost model. What it means GPT-5.6 Sol keeps the single-shot crown and the things that come with it: the highest first-try rate, the highest reliability, and by far the fastest and most concise runs. You pay roughly double for that, and you must guardrail its 20% regression rate. GLM-5.3 is the value and best-of-k pick: within four points on the first shot, ahead on pass@2, pass@4, and coverage, at half the price, with a cleaner failure profile and the best JavaScript on the board. Its real weaknesses against Sol are protocol conformance and raw speed. And because the two genuinely diverge, the sharpest deployment is neither alone. It is GLM-5.3 as a lower-cost front end with Sol as the verifier-gated escalation, which lands flagship-beating coverage for less than the flagship's own per-task price. DeepSWE · Full Results GLM 5.3 vs GPT 5.6 Sol, metric by metric Metric GLM 5.3 [max] GPT 5.6 Sol [max] pass@1 (official scoring) 69.0% 72.7% pass@1 (errors as failures) 68.8% 72.3% pass@2 / pass@4 81.1 / 87.6% 81.0 / 85.8% Coverage / reliability 87.6 / 78.8% 85.8 / 84.5% Solid (4/4) / walls (0/4) 48 / 14 61 / 16 Cost per rollout / total $3.99 / $1,806 $8.37 / $3,783 Solves per $100 17 9 Avg minutes / steps 35 / 124 19 / 61 Avg peak context / output tokens 155k / 80k 187k / 60k Failure anatomy (near miss / regression) 61% / 11% 54% / 20% Domains won (of 8) 4 4 Languages won 2 (JavaScript, Rust) 3 (Python, Go, TypeScript) Per-task correlation / union 0.43 / 106 of 113 (93.8%) Cascade GLM → Sol (accuracy / cost) 85.9% / $6.61 (vs Sol alone 72.7% / $8.37) Oracle 1-shot router 83.8% Infra errors 1 2 113 DeepSWE tasks · 4 trials per config · both at max effort · 904 rollouts total FAQs Is GLM-5.3 better than GPT-5.6 Sol? It depends on the metric. GPT-5.6 Sol wins single-attempt quality on DeepSWE (pass@1 72.7% against 69.0%), solves more tasks four for four (61 against 48), and is roughly twice as fast per rollout. GLM-5.3 ties pass@2 and wins pass@4 (87.6% against 85.8%) at half the cost per rollout, so it is the stronger value pick for high-volume or retry-tolerant agent work. How much cheaper is GLM-5.3 than GPT-5.6 Sol? In our run, GLM-5.3 cost \$3.99 per rollout against \$8.37 for GPT-5.6 Sol at max effort, about 2.1x lower. Measured per solved task, GLM-5.3 returned 17 solves per \$100 against Sol's 9, roughly twice the solved work per dollar. Which is better for coding, GLM-5.3 or GPT-5.6 Sol? They split the board. GLM-5.3 takes JavaScript (90 against 75) and Rust (70 against 60); Sol takes Python (74 against 66), Go (79 against 76), and TypeScript (66 against 61). By task domain it is four wins each: Sol leads the exact-contract and systems work, GLM-5.3 leads query and config languages, runtime internals, and stateful reactivity. Should I route between GLM-5.3 and GPT-5.6 Sol? Yes, if you can verify results. The two diverge (0.43 correlation) and fail differently, so running GLM-5.3 first and escalating to Sol when your test suite rejects the output reaches 85.9% at \$6.61 per task, beating Sol alone (72.7% at \$8.37) and a perfect one-shot oracle router (83.8%). Together the two cover 106 of 113 tasks. What is pass@k on DeepSWE? pass@k measures whether at least one of k attempts at a task passes the hidden test suite. pass@1 rewards getting it right first try; higher k rewards a model that can eventually reach a solution across several tries. GLM-5.3's edge grows as k increases. Method and caveats Data: DeepSWE v1.1 export, 113 tasks by 4 trials per config, both at max effort, from the published per-trial records. 452 trials per side. Scoring: headline pass rates use DeepSWE's official scoring (included_in_score, infra errors excluded). GLM-5.3 has 1 infra error and Sol 2, so official and strict scoring nearly coincide. pass@2, pass@4, coverage, union, and correlation use per-task pass counts. Costs are the published per-trial cost_usd from the index. Per-turn trajectory JSONs for the GLM-5.3 batch were not on the public CDN at analysis time, so this analysis is index-level. Failure anatomy uses published per-test fractions. A near miss is at least 80% of new tests passing with the baseline intact; a regression means a baseline test broke. Domains use an artifact-based task taxonomy. The cascade assumes a verifier decides escalation and independence given the task. Expected accuracy is symmetric in order, cost is not, so lead with the lower-cost model. The oracle router is the per-task best single-shot pick, an upper bound on any one-shot router.