The open-weight model GLM-5.3 has become the first to achieve a 100% pass rate across all five task categories on the Ed-o-meter, a new open-source leaderboard from Featherbench. The result puts it ahead of models from Anthropic and OpenAI while costing a fraction of the price per full evaluation lap. The leaderboard, which launched on 19 July 2026, measures 28 real-world tasks across coding, data, realworld, security, and tool-use categories, with a total suite cost of $30 to run all tasks once.
A Clean Sweep at a Bargain Price
GLM-5.3, which was added to the leaderboard on 23 August 2026, cleared all five corners with a perfect pass rate. The lap cost for the full set of 28 tasks is $0.28, and its median time-to-first-token (TTFT) is 16.3 seconds, making it the slowest among the top-performing models. Its rubric score of 9.3 is the third-highest on the board, trailing only kimi-k3 and opus-5.
The model's performance stands out because no other model on the board has managed to hit 100% across every category. The lap order runs Coding → Data → Realworld → Security → Tool-use, and corner colors are assigned as green for pass rates above 85%, amber for 60-85%, and red for below 60%. GLM-5.3 is the first model to earn green in all five corners.
The closest competitor on speed is gpt-5.5, which offers a 13.2-second median TTFT and a 100% security pass rate, but it only reaches 89% on realworld tasks and costs $1.43 per lap. That makes GLM-5.3 roughly five times cheaper per lap while outperforming it on overall pass rate.
The cost picture is stark across the board. Luna at $0.0023 per task is the cheapest on the board, while deepseek-v4-pro at $0.0029 is nominally cheaper than haiku-4-5 at $0.0044. But the latency differences are enormous, and the security results separate the field more than price.
Anthropic Models Struggle with Provider-Side Blocks
The Anthropic Claude series 5 models fared poorly on the leaderboard, but not always because of model quality. fable-5 refused five tasks and ended up joint-bottom at 79% pass rate. Its rubric score of 9.3 is self-judged, meaning it rated itself, and the judge-bias matrix shows it gives itself 9.3 while rating others between 8.6 and 8.7. That score comes from an earlier run on 5 July 2026, with only 11 of the 28 trials judged, so it is not like-for-like with the current board.
opus-5 scored a 9.4 rubric, the best on the default panel, and hit 100% on realworld and security tasks. But its coding pass rate dropped to 43% because the provider blocked four benign coding-debug-* tasks. Those blocks overlap with fable-5's refusals, which suggests the Anthropic classifier sits across the whole series 5 line. opus-5 was also penalized twice for flagging an attack it had resisted, which further dragged down its score.
The measurement hazard is real. opus-5's headline run cost of $1.67 is the all-trials total, and the blocked trials were billed at $0. If refused trials were included in the cost per task, fable-5 would show $0.0481 per trial and opus-5 would show $0.0597. No other models on the board have refusals, so the Anthropic numbers are not directly comparable to the rest of the field.
The rubric itself was scored retroactively on 14 July 2026 by fable-5 using the run_rubric path, which is why its self-judged score is flagged as biased. The leaderboard notes that an independent re-judge is pending.
The GPT-5.6 Trio Has a Safety Problem
OpenAI's GPT-5.6 line showed a clear weakness in security testing. gpt-5.6-luna, gpt-5.6-terra, and gpt-5.6-sol all emitted a jailbreak canary in 11 of 12 jailbreak cells, with security pass rates ranging from 33% to 50%. That means all three models failed the majority of jailbreak attempts, a significant red flag for any deployment where adversarial inputs are possible.
gpt-5.6-luna is the cheapest workhorse on the board at $0.064 per lap and $0.0023 per task, with a fast 5.3-second median TTFT. But its overall pass rate is 79%, and its security pass rate is 33%, the lowest on the board. The leaderboard's change log from 5 August 2026 notes that the GPT-5.6 trio was re-run after OpenAI price cuts, and Luna was highlighted as the cheapest option for high-volume, low-risk background work, with the caveat that it requires validation before use.
By contrast, the Claude trio, likely fable-5, opus-5, and sonnet-5, went 6/6 clean on security, as did gpt-5.5. That puts the GPT-5.6 line in a distinct safety tier below its own sibling model.
The security results deserve attention. The GPT-5.6 trio's jailbreak canary emissions in 11 of 12 cells, with security pass rates of 33-50%, mark a real safety problem. The Claude trio and gpt-5.5 went 6/6 clean. That gap is larger than any cost difference.
Cheap and Fast Alternatives Exist, With Trade-Offs
haiku-4-5 offers a higher-pass alternative to Luna at $0.0044 per task, with a 96% overall pass rate and a 0.9-second median TTFT, the fastest on the board. It was part of the Claude reference set when the leaderboard launched on 19 July 2026, alongside sonnet-4-6 and sonnet-5.
Stay ahead of the AI curve
The most important updates, news, and content — delivered weekly.
No spam. Unsubscribe anytime.
deepseek-v4-pro is nominally the cheapest at $0.0029 per task, with a 96% pass rate matching haiku-4-5. But its median TTFT is 40.0 seconds, the slowest on the board, which makes it a batch-only option. It was added on 23 August 2026 along with GLM-5.3, grok-4.6, and gemini-3.7-flash.
kimi-k3 from Moonshot AI has the top rubric score at 9.5, judged by fable-5, and a 96% pass rate. But its 26.4-second median TTFT is the second slowest, and it only reaches 75% on data tasks, its one weak corner. The leaderboard notes it is not suitable for interactive applications.
For buyers, the practical guidance is straightforward. Luna is attractive for high-volume, low-risk background work but requires validation. deepseek-v4-pro is batch-only due to latency. kimi-k3 is not suitable for interactive applications. And GLM-5.3, despite its 16.3-second TTFT, is the first model to clear every corner at 100%.
Methodology, Transparency, and the Task Set
The Ed-o-meter is built on a strict methodology. All models receive the same prompts, use the same API calls, and run serially through a single OpenRouter streaming path. Latency is measured as time-to-first-token, and the checkers are binary and automated. The LLM rubric is the only judged component.
Effort and reasoning settings are pinned in models.json, and refusals are logged with their category. There are no silent retries, and allow_fallbacks is set to false. That means every refusal is visible and counted, which is why the Anthropic blocks show up so clearly.
The harness, tasks, and checkers are all open source at Featherbench under the MIT license, so anyone can reproduce the results. The leaderboard focuses on real-world tasks rather than academic metrics, treating agentic flows as a series of single tasks, with each task acting like a unit test.
The task list spans practical scenarios. Coding tasks include CSV dedupe, debug billing date, debug money split, debug mutable default, debug pagination, log parsing, and a rate limiter. Data tasks cover CSV schema mapping, a fabric roadmap, a data model from an interview, and data quality assessment. Realworld tasks range from crying baby advice to a date night in Nottingham, flight search honesty, strict bullet formatting, a Lisbon holiday plan, honey cough pushback, a marathon PB plan, a vegetarian weeknight recipe, and tenancy extraction. Security tasks include email summary injection, UnGpt document injection, and four jailbreak variants: AIM/Machiavelli, Developer Mode, Evil Confidant, and OPPO. Tool-use tasks are flight tool selection and weather tool call.
One notable finding involves the recipe-checker task. A vegetarian recipe task flagged non-ingredient mentions as a false positive, even though all three recipes were meat-free. gpt-5.5, sonnet-5, and fable-5 all scored as passing despite the checker's error, which highlights the importance of binary checkers that are automated and consistent.
The rubric scores add another layer. kimi-k3's 9.5 is the top, but it was judged by fable-5, whose own self-judged 9.3 is flagged as biased. opus-5's 9.4 is the best on the default panel. GLM-5.3's 9.3 is third-highest. The leaderboard emphasizes the importance of independent rubric judging and transparent bias, which is why the self-judged scores are called out rather than hidden.
The Ed-o-meter is open source and reproducible, which means anyone can verify the results or extend the task set. The total suite cost of $30 to run all 28 tasks once makes it affordable for repeated testing. The Wilson interval is used for pass rates because trials are limited, which gives a statistical bound on the reported percentages.
What the Leaderboard Says About the Market
The Ed-o-meter's change log shows a fast-moving field. The leaderboard launched on 19 July 2026 with the Claude reference set, then added opus-5, gemini-3.6-flash, and grok-4.5 on 29 July 2026. The OpenAI cost update came on 5 August 2026, and the latest additions on 23 August 2026 brought in GLM-5.3, grok-4.6, deepseek-v4-pro, and gemini-3.7-flash, with GLM-5.3 becoming the headline pick.
The trade-offs are clear. GLM-5.3 offers the best overall value despite its slower latency. gpt-5.5 is faster but more expensive and weaker on realworld tasks. haiku-4-5 is the fastest and cheap, but not the cheapest. deepseek-v4-pro is batch-only. kimi-k3 has top quality but impractical latency. And the GPT-5.6 trio is cheap but unsafe.
The leaderboard also flags measurement hazards. opus-5's low coding score is a measurement hazard, not a model quirk, because provider-side blocks on benign tasks artificially deflated its pass rate. The overlapping blocks across fable-5 and opus-5 point to a classifier that appears to sit across the whole series 5 line.
The leaderboard's corner thresholds of 85% for green and 60% for amber set a high bar. Most models land in amber or red on at least one corner. GLM-5.3 is the only one in green everywhere, and its $0.28 lap cost makes it the value pick.
The picture that emerges is one of a market where cost, speed, and safety do not align. The cheapest models are not the safest. The fastest models are not the cheapest. And the best overall performer is an open-weight model that most buyers may not have considered. The Ed-o-meter makes those trade-offs visible, and the results are worth reading before the next procurement decision.