What changed on llm.buzz

A sampling setting a model refuses is now dropped and named, instead of failing the request.

• The Claude models that only think adaptively (Opus 4.7 onward, Sonnet 5, Haiku 5, Fable 5, Mythos 5) take no temperature, top_p or top_k. A value you send is dropped and named in x-sgr-ignored-parameters as temperature->dropped(unsupported_by_provider). Haiku 5 now thinks adaptively, like the rest of the 5 generation.
• OpenAI’s reasoning models take temperature and top_p only where OpenAI does. GPT-5, GPT-5 mini, GPT-5 nano, GPT-6 Astra and GPT-6.1 Sol take neither; the others take them only at reasoning effort none, and otherwise drop them as temperature->dropped(set_reasoning_effort_none).
• GPT-5.1 and GPT-5.2 now run at reasoning effort none when you name no effort, like the GPT-5.4 family, so a temperature with no effort reaches them.
• The reasoning efforts for GPT-5.6 and GPT-6 match OpenAI’s current values: the GPT-5.6 family, GPT-6 Luna and GPT-6 Sol take none through max without minimal, and GPT-6.1 Sol takes low through max.
• For providers: the capability check sends no sampling settings, and on Claude models that refuse a forced tool choice its tools step lets the model pick the tool.

Read more: Sampling settings a model does not take, Dropped fields are not errors

A provider can now run a discount for a set time, from 1 to 99 percent off every price it bills.

• While it runs, every request that provider serves is billed at the discounted price, including a request that started before the discount ended.
• smart and cheapest routing compare the discounted prices.
• The model page shows a shield and “N% off” next to that provider, with the original price struck through beside the price you pay.
• GET https://api.llm.buzz/api/v1/models/{author}/{slug}/endpoints lists the discounted prices with discount, the fraction off (0.51 for 51% off), as OpenRouter does, and 0 when none runs.
• The price history on a model page shows prices before discounts.

Read more: Discounts

Each provider’s prices now sync from its own published price list, such as DeepInfra’s, Novita’s or Scaleway’s, so the price you pay follows the price the provider publishes.

• Scaleway’s prices come from its public product catalog. Scaleway lists them in euros, converted to US dollars at the European Central Bank’s daily rate plus 3%, the cost of converting. Every price you see or pay is in US dollars.
• Where a provider publishes no list, or a price is set by hand, that price is what you pay and what the model pages show.

Read more: Which providers serve

New models now start on the smart routing preset, which orders providers by price and time to first token together.

• Every provider of a model passes the automated capability check before it serves, so each one serves every capability the model declares: a streamed answer with usage, then tool calls, structured output and image input where the model declares them.
• On a model page, the answer to “How much does this model cost?” gives the price of the provider requests go to first, and names that provider.

Read more: How a provider is chosen, Which providers serve

The llm.buzz header now searches the whole site, and the Models list and model pages show more of each model at a glance.

• Search from the header, or press Cmd+K, Ctrl+K or /. It finds models, providers, pages and docs; “See all” opens Models filtered by your query.
• Each row on Models shows up to three lines of the model’s description, its input and output modalities as colored icons, and its features (open weights, tools, reasoning, caching, structured outputs) as icons with a tooltip. A button copies the model ID.
• A model page shows its whole description, with the same modality and feature icons.
• Compare picks and swaps models with the same search.
• The latency and throughput charts on model pages cover the last 24 hours, hour by hour, drawn as one continuous line through the measured hours.
• A supplier who signs in on the supplier sign-in page with an account that already exists is taken to that account.

A request can now run for 600 seconds, up from 120, and a stream cut at the limit ends cleanly and is billed only for what you received.

• A stream still generating at the limit ends with a timeout error chunk and the stream terminator, and is billed for the output delivered, not for its whole reservation. Any other request past the limit answers 504 deadline_exceeded.
• When web search is busy, the search is skipped instead of holding up the request. The answer still arrives, no search is billed, and x-sgr-ignored-parameters names it as web_search->dropped(search_busy).

Read more: Time limit, Errors