OpenRouter is Quietly Giving Nerfed AI Model (STOP Using Wrong Endpoint)

OpenRouter publishes the precision of almost every endpoint it sells, and the field everyone is reading turns out to be the wrong one. Pull OpenRouter's own endpoints API and you get 335 endpoints across 54 providers for 27 open-weight models. 54 of them serve something at four bits and 120 of them, 35.8%, do not say what they serve at all. That looks like the whole story until you open the labs' own config files: Kimi K3 ships as MXFP4 with quantization-aware training, Kimi K2.6 ships as INT4 in groups of 32, OpenAI's gpt-oss-120b stores its expert weights in MXFP4 and ran all of its own published evals that way, and DeepSeek V4 Pro has a field in its config that reads expert_dtype: fp4. Moonshot's own endpoint is one of the four-bit providers. An "fp4" tag is usually a provider telling you it changed nothing, and a "bf16" tag on a natively four-bit model is an upcast that cannot add back information that was never stored. Artificial Analysis settled it in August with a per-endpoint accuracy index rather than a per-model one. On GLM-5.2, Nebius scored 100% and DeepInfra scored 73%, and both of them are tagged fp4. The most famous provider gap there is, gpt-oss-120b at 80% on AIME 2025 at Azure against 93.3% at six other providers, turned out to be old vLLM commits that ignored reasoning_effort, and it was fixed inside a day. What actually varies is the room each endpoint gives the model. Novita will serve Llama 3.3 70B with 12,288 tokens of context where Meta ships 131,072. Together caps the answer at 2,048 tokens. DigitalOcean's gpt-oss-120b caps output at 4,096, and Epoch AI found the model needs more than 32,000 to finish a hard benchmark question. Across all 335 endpoints, 27 advertise under half the model's context, 19 cap output at 16,384 or less, 26 cannot call tools, and 121 ignore the seed you send, because require_parameters defaults to false and the docs say unsupported parameters are silently ignored. The real cut is in the price sheet. OpenRouter's documented default takes the lowest-cost candidates and picks one weighted by the inverse square of the price, and the docs never say which price. GLM-5.3 sells at 3 cents in and 16 dollars out on one endpoint and 42 cents in and $1.32 out on another, so output-to-input runs 3.1x at Novita, 3.1x at Z.AI's own endpoint, and 533x at Relace. Price the same traffic at six tokens read per token written and the ranking inverts: Novita costs 55 cents per million, Relace $2.31, which is 4.2 times the price of the endpoint the router favours least. Cached input, the thing an agent pays for on every turn, is read by the sort not at all. WHAT THIS VIDEO COVERS "fp4" on OpenRouter is usually the lab's own format, and four of the nine models checked ship natively at four bits Two endpoints at the same advertised precision scored 27 accuracy points apart, so the precision label is not the variable The cheapest Kimi K3 endpoint stores more bits than Moonshot's own and charges 39% less for output The silent cuts that do show in the data are context windows, output caps and ignored parameters, not weights OpenRouter's own Auto Exacto numbers say its default route was failing about 8% of GLM-5 tool calls before March, and the fix only covers requests carrying tools The four lines of JSON that turn the routing lottery off, and the one that accidentally opts you back out of the quality router CHAPTERS 00:00 The 533x price sheet 01:16 Why quantization is the obvious suspect 02:36 The labs ship four bits themselves 04:00 The cheapest Kimi has more bits 05:16 Two fp4 endpoints, 27 points apart 06:36 What varies is the room 08:08 Where the cut is real 09:20 OpenRouter fixed half in March 10:44 The square that picks for you 12:12 What your invoice actually pays 13:40 Who should ignore all this 14:52 Four lines of JSON SOURCES OpenRouter provider routing docs (inverse square of price, provider object, quantization enum) https://openrouter.ai/docs/guides/routing/provider-selection OpenRouter model endpoints API (live capture 2026-10-05, 335 endpoints / 54 providers / 27 models) https://openrouter.ai/api/v1/models OpenRouter, "Provider Variance: Introducing Exacto" (2025-10-21) https://openrouter.ai/blog/announcements/provider-variance-introducing-exacto/ OpenRouter, "Auto Exacto: Adaptive Quality Routing, On by Default" (2026-03-12) https://openrouter.ai/blog/announcements/auto-exacto/