ERNIE 5.5 Flash Preview: API Pricing and Agent Setup

ERNIE 5.5 Flash Preview is Baidu’s first post-trained release built on ERNIE 5.5 Flash Base. It is a text-in, text-out model with 280B total parameters, 14B active per token and a 1M-token context. Baidu positions it for coding, multi-step agent work, knowledge tasks and content creation. The launch shows publisher-selected game, 3D and document projects built inside an agent framework.

Preview and Base serve different jobs. Preview is the hosted, post-trained model available through Qianfan’s API; Base is the earlier pre-training checkpoint for people who want to adapt or serve weights. Baidu says Preview weights will be released under Apache 2.0, but its announcement still says “soon.” Use the hosted Preview for ready-to-prompt tasks and the separately licensed Base checkpoint for weight adaptation.

Baidu describes Preview as an intermediate post-training checkpoint and says training is still in progress. Treat it as a candidate to evaluate on representative tasks, especially if your workflow depends on consistent agent behavior.

Qianfan lists ernie-5.5-flash-preview as supporting function calling. You provide a function schema; the model can return a proposed call and arguments, but your application must validate them, run the function and send its result back. Baidu estimates that the built-in function-call prompt adds about 300 input tokens, so include that overhead when you estimate a session.

Qianfan’s native web_search option is not supported for this model. If your agent needs current web results, connect a search tool through your own function loop or choose an endpoint with built-in search. A one-million-token context can hold long project material, but the tokens you actually send, including history and tool results, are what affect the bill.

The Qianfan API does not support temperature, top_p or seed for this model. Thinking cannot be disabled or tuned; the documented reasoning setting is fixed at high. The max_tokens limit counts reasoning and the final answer together. Set the output budget for both reasoning and the final answer.

Check the output caps before long generations: Qianfan’s product page lists 64,000 tokens as the default output for Preview, while the model table lists 1M-token maximums for input and output. A single request cannot spend both maximums at once.

GameFeel Benchmark v1.5 records Ernie-5.5 Flash Preview at 43.6 Overall, with its effort column labeled Max. Overall averages seven completed categories covering creation, bug fixing, game feel, ideas, visuals, audio and reverse engineering. This is the evaluator’s game-development score, rather than a general agent ranking or a Qianfan task-cost measurement.

Use those task categories to build a small test around your own game: ask for one mechanic, a reproducible bug fix and a visual or audio revision. Inspect the playable result and count every repair. The public dataset’s output-speed field does not include enough configuration detail to estimate your complete session time or bill.

Baidu’s Qianfan price page lists a 50% promotion through its listed November 30, 2026 deadline. The table below shows the promotional rate and the regular rate it returns to afterward. Rates are yuan per million tokens; cached input is billed at the cache-hit rate only when the request actually receives a cache hit.

For one call using 100,000 uncached input tokens and 20,000 output tokens, the promotional bill is ¥0.18 in the listed peak window: 0.1 × ¥1 input + 0.02 × ¥4 output. At regular rates, the same usage is ¥0.36. In the listed idle window the corresponding totals are ¥0.09 and ¥0.18. This is billing arithmetic for fixed token counts, not a quality or speed comparison. Each agent turn, tool result, repeated conversation history and retry can add more usage; inspect the returned usage fields for a real run.

Qianfan switches between peak and idle rates on its billing schedule. The linked public table leaves the timezone unspecified; use the schedule shown in your account when timing a batch.

OpenRouter lists baidu/ernie-5.5-flash-preview:free as a free route with rate limits. It is a convenient way to test a prompt without setting up Qianfan billing, but its availability, limits and provider routing are separate from Baidu’s Qianfan price schedule. Recheck the live listing before building a service around a free tier.