ModelPort

LLM upgrade errors and fixes: Claude 5.x, GPT-6, Gemini 3.8

Symptom to cause to fix, across providers. Use it when a model swap starts returning errors or behaving differently. Each row points at the guide that has the full migration. Facts verified 2026-10-11 against provider docs (see each guide’s Source: line). Exact error text is quoted where the provider documents it. Where it is not documented here, the symptom is described instead of guessing a string.

Request errors (HTTP 400)

You send Fails on Error or symptom Fix
thinking: {type: "disabled"} Claude Opus 5.5, Fable 5.1, Fable 5 "thinking.type.disabled" is not supported for this model. Use "thinking.type.adaptive" and "output_config.effort" to control thinking behavior. Omit thinking. Set output_config.effort to low for speed.
thinking: {type: "disabled"} Claude Sonnet 5.5 "thinking.type.disabled" is not supported for this model. Use "thinking.type.between_tools" for the lowest thinking setting, or ... Try adaptive + effort: "low". Else {type: "between_tools"} at effort high or below.
thinking: {type: "enabled", budget_tokens: N} Claude Opus 4.7 and later, Sonnet 5 and later, Haiku 5.5, Fable 400 {type: "adaptive"} + output_config.effort.
thinking: {type: "disabled"} with effort xhigh or max Claude Opus 5, Haiku 5.5 output_config.effort 'xhigh' is not supported when thinking is disabled on this model. Use effort 'high' or below, or enable thinking. Lower effort or enable thinking.
thinking: {type: "between_tools"} Any model except Claude Sonnet 5.5 "thinking.type.between_tools" is not supported for this model. Drop the field when re-sending to another model.
tool_choice: {type: "any"} or {type: "tool", ...} Claude Opus 5.5, Sonnet 5.5, Fable 5.1 tool_choice: type "tool" and "any" are not supported for this model. auto + strict: true + prompt instruction, or output_config.format for JSON. Check a call happened.
computer_20251124 tool Claude Opus 5.5, Sonnet 5.5 on Claude API and Google Cloud 'claude-opus-5-5' does not support tool types: computer_20251124. {"type": "computer_toolset_20260801"}. Amazon Bedrock still accepts the old tool.
computer_20250124 tool Claude Haiku 5.5 on Claude API and Google Cloud 400 computer_toolset_20260801.
temperature, top_p, top_k (non-default) Claude Opus 4.7 and later, Sonnet 5 and later, Haiku 5.5, Fable 400 Omit them. Haiku 5.5 accepts only temperature 1 and top_p 0.99.
Assistant prefill (last message is assistant) Claude Opus 4.6 and later, Sonnet 4.6 and later, Haiku 5.5, Fable 400 End with a user turn. Use output_config.format or a prompt instruction.
Replayed thinking block after editing system, tools, or earlier messages Claude Fable 5.1, Opus 5.5, Sonnet 5.5, Haiku 5.5 (accounts created on or after 2026-08-31) 400 Keep history append-only. See claude-fable-5-1.md.
fallbacks on Haiku 5.5 Claude Haiku 5.5 A model list returns 400. "default" leaves a decline declined. Handle the refusal in the client.
thinking_budget and thinking_level together Gemini 3.x 400 Send only thinking_level.
thinking_level: "minimal" Gemini 3.8 Flash Error Use low, medium, or high.
reasoning.effort: "none" or "minimal" GPT-6 Astra, GPT-6.1 Sol Unsupported Use low.
temperature, top_p, top_logprobs, logprobs with reasoning on GPT-6 family Unsupported Remove them.
Tool calling on Chat Completions GPT-6 Astra, 6.1 Sol Unsupported Use the Responses API.

Silent changes (no error, output differs)

Change Models Effect Fix
Thinking on by default Claude Opus 5, Sonnet 5, Haiku 5.5 max_tokens shared with thinking. Replies truncate. content[0] may be a thinking block. Raise max_tokens (64K for long agent turns). Read blocks by type.
Default effort medium Claude Opus 5.5, Haiku 5.5 Less thinking than the old high default. Set effort explicitly. Re-run the sweep.
thinking.display default "omitted" Claude Opus 4.7 and later, Sonnet 5 and later, Fable Thinking text is empty. UI shows a long pause. display: "summarized".
Progress notes moved into thinking blocks Claude Opus 5.5, Sonnet 5.5, Fable 5.1 UI goes quiet between tool calls. display: "updates" (beta) and render non-empty thinking blocks.
New tokenizer, about 30% more tokens Claude Sonnet 5 and later (vs 4.6), Haiku 5.5 (vs 4.5) Cost and context budgets shift. Recount with count_tokens.
Classifier refusals (HTTP 200, stop_reason: "refusal") Claude Fable 5.1, Opus 5 and 5.5, Sonnet 5.5, Haiku 5.5 Code reading content[0] breaks. Branch on stop_reason. Add a fallback where supported.
More literal instruction following Claude Opus 4.7 and later Vague prompts behave differently. State format, verbosity, and scope.
More tool-call iteration, smaller reasoning steps Gemini 3.8 Flash More tokens on long tasks. Lower thinking_level.
More clarifying questions, long formatted answers GPT-6 Astra Stalls on autonomous work, verbose output. Prompt for action and style. See gpt-6.md.

Where to go next

This site is open source. Improve this page.