Open-weight GLM-5.3 clears all 28 Ed-o-meter tasks at a fifth of GPT-5.5's cost

Open-weight model tops independent leaderboard

An independent benchmark that surfaced on Hacker News' front page on 23 August 2026 reports that GLM-5.3, an open-weight model, passed every one of its 28 real-world tasks — the first model on the board to manage it — while spending roughly a fifth of what OpenAI's gpt-5.5 needs for the same workload. According to the Ed-o-meter leaderboard, maintained by Ed Yau, an applied AI architect at Kerv, GLM-5.3 completed the full suite for $0.28 against gpt-5.5's $1.43, and posted 9.3 on the site's quality rubric, the third-highest among the 17 models tested.

One harness, deterministic grading

The Ed-o-meter runs every model through the same 28 tasks with identical prompts and API calls, routed through a single OpenRouter streaming path and executed one after another. Grading is binary and automated, except for a rubric component judged by an LLM, whose bias the site says it documents rather than assumes away. Tasks span five categories — coding, data, real-world work, security and tool use — and the entire suite costs about $30 to run. The harness, tasks and checkers are open source under an MIT licence as Featherbench, so anyone can rerun the suite and verify the numbers.

GLM-5.3's penalty is speed: a median 16.3 seconds to first token, against 13.2 seconds for gpt-5.5. The OpenAI model, however, reached only 89% on the real-world category while keeping a clean security record.

Safety filters masquerading as failures

Two Anthropic-family models posted scores the site attributes to measurement artifacts rather than capability. Opus-5 recorded the best rubric on the default panel at 9.4, with 100% on real-world and security tasks, but just 43% on coding — because, per the site, a provider-side classifier blocked four benign debugging tasks before a single token was generated. Fable-5 hit an overlapping wall, refusing five tasks outright and finishing joint-bottom at 79%. Yau treats this as a hazard running across the series-5 line rather than a quirk of either model, and notes opus-5 was twice penalised for flagging an attack it had actually resisted.

The security category produced the sharpest split on the board. The three gpt-5.6 variants produced a jailbreak canary in 11 of 12 jailbreak test cells, for security pass rates of 33–50%. The Claude trio and gpt-5.5 came through clean.

The budget end of the board

For high-volume background work, the site picks gpt-5.6-luna at $0.0023 per task with a 5.3-second median time to first token — but flags its 79% overall pass rate and 33% security score as reasons to validate every output and keep it away from untrusted prompts. Haiku-4-5 is the higher-accuracy alternative at $0.0044 per task, 96% overall and a 0.9-second response. Deepseek-v4-pro matches that 96% at $0.0029 per task but takes a median 40 seconds to first token, the slowest on the board, which the site says rules it out for anything interactive. Kimi-k3 tops the rubric at 9.5 with 96% overall, but a 26.4-second median latency and a 75% showing on data tasks confine it to non-interactive use.

Caveats on the numbers

The results come from a single trial per model over 28 tasks, and the site itself reports wide Wilson confidence intervals as a result. Effort and reasoning settings are pinned and published alongside every number, refusals are logged rather than silently retried on another model, and routing is locked so no quiet fallbacks to quantized variants can occur.

Why it matters

If the result holds up beyond a single trial, an open-weight model matching or beating frontier closed models on practical tasks at a fifth of the running cost changes the economics of model selection and strengthens the case for self-hosting. The run also carries two broader lessons for anyone reading LLM leaderboards: provider-side safety filters can crater a benchmark score without saying anything about model ability, and the cheapest options can carry genuine security exposure. One operator's 28-task lap is a useful signal — not a verdict.