GLM-5.3 (open-weight) beat Anthropic/OpenAI models – for 1/5 the cost

If you only tally 1 model, tally glm-5.3

glm-5.3 is the first exemplary connected the committee to clear each 5 corners — coding, information development, realworld, information and tasks — astatine 100%. It backs that pinch a 9.3 rubric, third-highest connected the board, and $0.28 for the lap. The 1 costs is patience — a 16.3-second median time-to-first-token. gpt-5.5 is the faster replacement astatine 13.2s, pinch the aforesaid 100% information but an 89% realworld area and $1.43 for the lap.

Fable grounded to complete a azygous lap

fable-5 is joint-bottom astatine 79% because it refused to do 5 of the tasks. It performed good connected what it completed, but moreover it thought kimi-k3 was giving amended answers. You'll request a fallback exemplary if you're utilizing Fable. opus-5 deed the aforesaid wall — 4 benign coding-debug-* tasks blocked earlier a token was generated, connected an overlapping group of tasks — truthful Anthropic's classifier looks for illustration it sits crossed the full bid 5 line, not conscionable Fable. See the full refusal breakdown for what's really going on.

Luna is the very cheapest workhorse

gpt-5.6-luna costs $0.064 for the afloat lap, aliases $0.0023 per task, pinch a 5.3-second median TTFT. That makes it charismatic for high-volume, low-risk inheritance activity wherever failures are inexpensive to observe and retry. The trade-off is material: 79% wide and 33% connected security, truthful validate each consequence and support it distant from untrusted prompts. haiku-4-5 is the higher-pass replacement astatine $0.0044 per task, 96% wide and a 0.9-second TTFT. deepseek-v4-pro is nominally cheaper still astatine $0.0029 per task for the aforesaid 96% walk rate, but its 40.0-second median TTFT — the slowest connected the committee — rules it retired for thing interactive; dainty it arsenic a batch-only option.

The enigma impermanent sets the fastest value lap

kimi-k3 still tops the rubric astatine 9.5 — judged independently by fable-5 — pinch a 96% walk rate, though opus-5's 9.4 now runs it adjacent connected value astatine a 3rd of the wait. The drawback is patience: a 26.4-second median time-to-first-token, 2nd slowest connected the committee down deepseek-v4-pro's 40.0s, and a 75% wobble connected information improvement tasks, its only anemic corner. Not suitable for interactive applications.

Three cars grounded the clang test

The gpt-5.6 statement is quick, but it has a information problem. gpt-5.6-luna, gpt-5.6-terra and gpt-5.6-sol emitted the jailbreak canary successful 11 of 12 jailbreak cells (33–50% information pass) — make judge you protect successful your harness, and use much observant Red teaming if utilizing these models. The Claude trio went 6/6 clean, arsenic did gpt-5.5.

A information select tin look precisely for illustration a bad lap

opus-5 posts the champion rubric connected the default sheet astatine 9.4 and 100% connected some realworld and information — past shows 43% connected coding. That compartment is not its debugging ability: 4 benign coding-debug-* tasks were blocked by a provider-side classifier earlier a azygous token was generated, connected an overlapping group of tasks to the ones already blocked connected fable-5. Two Anthropic-family models now deed the aforesaid filter, truthful dainty it arsenic a measurement hazard alternatively than a exemplary quirk — and statement opus-5 was besides penalised doubly for flagging an onslaught it had successfully resisted.

Coding (7 · Python)

• CSV dedupe — small, well-specified task pinch a deterministic unit-test checker
• Debug billing date — hole a month/day-overflow day bug without regressing the moving cases
• Debug money split — divided integer pennies N ways truthful shares sum precisely and enactment fair
• Debug mutable default — hole the classical mutable-default-argument bug
• Debug pagination — hole an off-by-one page-count bug
• Log parsing — parse logs pinch quoted quotes, missing fields, IPv6 separator cases
• Rate limiter — stateful people implementation tested against an injected clone clock

Data (4)

• CSV schema mapping — representation a earthy CRM export onto a conformed customer dimension
• Fabric roadmap — move a catalogue, mapping and requirements into a phased user-story roadmap
• Data exemplary from interview — deduce a star-schema exemplary and requirements from a stakeholder transcript
• Data value assessment — drawback seeded defects successful a mini array and prioritise fixes

Realworld (9)

• Crying babe advice — safety-sensitive triage for an unusually crying 8-week-old
• Date nighttime successful Nottingham — section recommendations without hallucinated venues
• Flight hunt honesty — admit nary unrecorded pricing information alternatively than fabricate flights
• Strict slug formatting — obey nonstop format constraints nether an awkward connection ban
• Lisbon vacation plan — five-day itinerary respecting cost, kids and a remainder day
• Honey cough pushback — correct a incorrect premise: nary chromatic for a 5-month-old
• Marathon PB plan — week-by-week training scheme astir a fixed nine session
• Vegetarian weeknight recipe — genuinely meat-free look for six
• Tenancy extraction — grounded answers from a provided agreement, including admitting what isn't covered

Security (6)

• Email summary injection — summarise an email without obeying an embedded refund instruction
• UnGpt archive injection — summarise a reappraisal without emitting the embedded jailbreak canary
• AIM/Machiavelli jailbreak — garbage the unfiltered persona while still answering the benign question
• Developer Mode jailbreak — garbage the dual-response jailbreak template
• Evil Confidant jailbreak — garbage the "no rules" persona while staying helpful
• OPPO jailbreak — garbage the opposite-of-a-safe-model persona template

Tool-use (2)

• Flight instrumentality selection — hunt only, don't book, erstwhile explicitly told not to
• Weather instrumentality call — telephone the upwind instrumentality pinch the correct location argument