Sampling Controls

Today StreamOptions exposes only Temperature and MaxTokens. Most other sampling knobs are unwired even where the provider supports them.

Core Sampling Parameters

Parameter	Anthropic	OpenAI Chat	OpenAI Responses	Google	pi-go option
temperature	✅	✅	✅	✅	✅ (options.go)
max_tokens	✅	✅ max_completion_tokens	✅ max_output_tokens	✅ maxOutputTokens	✅
top_p	❌	✅	✅	✅	❌
top_k	❌	❌	❌	✅	❌
frequency_penalty	❌	✅	✅	❌	❌
presence_penalty	❌	✅	✅	❌	❌

OpenAI reasoning models lock temperature; pi-go gates this through compat.SupportsTemperature (openai.go:304-305). MaxTokens is forced upward when thinking is enabled on Anthropic (anthropic.go:256) — caller’s request is silently overridden.

Stop Sequences

Provider	API	pi-go
Anthropic	✅ stop_sequences	❌
OpenAI Chat	✅ stop	❌
OpenAI Responses	✅	❌
Google Gemini	✅ stopSequences	❌

Seed / Reproducibility

Provider	API	pi-go
Anthropic	❌	—
OpenAI Chat	✅ seed + returns system_fingerprint	❌
OpenAI Responses	✅	❌
Google Gemini	❌	—

Logprobs

Provider	API	pi-go
Anthropic	❌	—
OpenAI Chat	✅ logprobs + top_logprobs	❌
OpenAI Responses	✅	❌
Google Gemini	⚠️ responseLogprobs (preview)	❌

Safety / Moderation

Provider	API	pi-go
Anthropic	⚠️ policy-driven; no per-call thresholds	❌
OpenAI Chat	⚠️ built-in filter; separate Moderation API	❌
OpenAI Responses	⚠️ same	❌
Google Gemini	✅ safetySettings per harm category (HARASSMENT, HATE_SPEECH, SEXUALLY_EXPLICIT, DANGEROUS_CONTENT, CIVIC_INTEGRITY) with thresholds	❌

Provider Documentation

• Anthropic — Messages params
• OpenAI — Chat completions params
• OpenAI — Responses params
• OpenAI — Reproducible outputs (seed)
• OpenAI — Logprobs
• OpenAI — Moderation API
• Google Gemini — generationConfig
• Google Gemini — Safety settings

pi-go Gaps

• top_p, top_k, frequency_penalty, presence_penalty not exposed for any provider that supports them.
• Stop sequences missing entirely; StopReason has no value distinguishing “matched stop sequence” from regular stop.
• seed missing; no SystemFingerprint on returned messages.
• Logprobs: no content type or message field for per-token logprobs. Adding requires extending ai.Text or a parallel metadata field.
• Safety settings: no SafetySettings option; no way to surface a safety_ratings / blocked-output reason from Gemini responses; StopReason has no safety value.
• MaxTokens override when thinking is on (Anthropic) is silent — caller’s intent isn’t preserved.