ERNIE 5.5’s comparison with GPT-5.6 has a pricing catch: GPT-5.6 is retired

Baidu’s ERNIE 5.5 Flash Preview offers an unusually large single-call output allowance, but comparing it with OpenAI’s GPT-5.6 Luna risks overlooking the newer, cheaper model that replaced it. OrcaRouter’s analysis gives developers a useful reminder: a model comparison is only as current as the alternatives on its rate card.

Watch Desk analysis

What happened

OrcaRouter compares ERNIE 5.5 Flash Preview with GPT-5.6 Luna, noting that the latter was marked deprecated after GPT-6 Luna arrived. It lists GPT-5.6 at $0.20 per million input tokens and $1.20 per million output tokens, against ERNIE’s $0.28 and $1.13. The article says GPT-6 Luna is listed at $0.10 input and $0.50 output, on the same context size as its predecessor.

The comparison’s standout specification is ERNIE’s stated output ceiling of 921,600 tokens in one call. But OrcaRouter says there is no shared benchmark: ERNIE’s results come from Baidu’s own runs, while GPT-5.6’s cited results come from Artificial Analysis. Those figures cannot be treated as a like-for-like contest. Read OrcaRouter’s comparison.

Why it matters

For a developer pricing a workload, the practical comparison is not only ERNIE against an older Luna model. GPT-6 Luna’s listed rates are lower, so the earlier model’s prices may not reflect the choice available now. And a very large output allowance matters chiefly for work that genuinely needs a single response of that scale.

OrcaRouter’s comparison is useful as a guide to what is listed and where the evidence differs, not as a neutral head-to-head verdict. Its central caution is sound: pricing, context limits and benchmark scores answer different questions, especially when the scores come from different evaluators.

Our read

Check the current model identifier and pricing before building around a comparison table. ERNIE’s output ceiling is a concrete reason to investigate it for unusually long single-call tasks; it is not proof that it will do those tasks better or more cheaply overall. A benchmark assembled by two different referees is not a fair race, however impressive the scoreboard looks.

What to watch

• Whether Baidu publishes independent results or more detail on the cost of completed tasks.
• Whether the listed ERNIE and GPT-6 Luna prices or specifications change.
• Whether developers report results on workloads that need exceptionally long outputs.

Discussion spark: Would you consider ERNIE 5.5 for its very large output ceiling without an independent benchmark, or wait for comparable testing even if that means giving up the capability for now?

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.