feat(provider): add adaptive thinking and 1M context support for Claude Opus 4.6

[okhsunrog]

Adaptive thinking

• Add adaptive-thinking-2026-01-28 beta header for Anthropic provider
• Detect Opus 4.6 models and use adaptive thinking with effort parameter
• Support all effort levels: low, medium, high, max
• Older models continue to use manual thinking with budgetTokens

Opus 4.6 uses adaptive thinking where the model decides how much to think based on task complexity, guided by the effort parameter. This is more efficient than fixed budgetTokens as simple tasks use minimal thinking.

1M context window

• Add context-1m-2025-08-07 beta header for Anthropic provider (API key users)
• Fix compaction logic: when model.limit.input is set, compare input tokens only against that limit instead of the combined total (input + cache + output)

Without the beta header, Opus 4.6 enforces a 200k input token limit (the context window is still 1M — output/thinking tokens aren't affected). The opencode (OAuth) provider doesn't support the 1M beta header, so it relies on model.limit.input in models.dev to trigger compaction at the right time: anomalyco/models.dev#819

AI SDK upgrade

• Upgrade ai v5 → v6, @ai-sdk/anthropic v2 → v3, and all @ai-sdk/* packages
• Migrate LanguageModelV2 → LanguageModelV3, async toModelMessages, renamed tool factories

Ref: https://platform.claude.com/docs/en/build-with-claude/adaptive-thinking
Ref: https://platform.claude.com/docs/en/build-with-claude/context-windows#1-m-token-context-window

Tested locally, works with Opus 4.6

Closes #12323
Closes #12338
Closes #12438

[github-actions]

Thanks for your contribution!

This PR doesn't have a linked issue. All PRs must reference an existing issue.

Please:

1. Open an issue describing the bug/feature (if one doesn't exist)
2. Add Fixes #<number> or Closes #<number> to this PR description

See CONTRIBUTING.md for details.

[github-actions]

The following comment was made by an LLM, it may be inaccurate:

No duplicate PRs found

[okhsunrog]

Adding of adaptive thinking needs an upgrade of @ai-sdk/anthropic which leads to upgrading of a bunch of other packages. Making it a draft for now.
I started working on upgrading the deps, I hope it can be made as part of this PR

[okhsunrog]

Added a commit with dependencies upgrade. Tested locally, switching effort levels with Opus 4.6 now works as expected

[okhsunrog]

Can anyone test if 1M context for you? It works for me just fine with Opus 4.6. I managed to get up to 340k token

[ItsWendell]

@okhsunrog what's the easiest way to test this?

[okhsunrog]

@ItsWendell make sure you have latest bun installed, then run these commands:

git clone -b feat/opus-4-6-adaptive-thinking https://github.com/okhsunrog/opencode.git
cd opencode
bun install
bun dev

[ItsWendell]

Tested and works:

[ItsWendell]

Couple of issues that I ran into, at e.g. 400 tokens I again get a error:

[okhsunrog]

well, shit...

[okhsunrog]

@ItsWendell yes, I confirm the issue is there. but if we'd pass context-1m-2025-08-07 header we'd get error from Anthropic with OAuth, that the long context beta is not available for this subscription. Are you using it with Claude subscription via OAuth or vie Anthropic API?

[okhsunrog]

Confirmed, the 200k limit is enforced by the Anthropic API. The status bar shows a combined total (input + output + cache tokens), but the API only counts input tokens against the 200k limit. That's why I didn't hit it earlier — actual input tokens stayed under 200k even though the status bar showed 385k.
The context-1m-2025-08-07 beta header is needed for Opus 4.6 to go past 200k, same as Sonnet 4/4.5. The compaction check in opencode uses this combined count:

// compaction.ts:35
const count = input.tokens.input + input.tokens.cache.read + input.tokens.output

So opencode thinks there's plenty of headroom (39% of 1M) while the API is already at the 200k input limit. I'll update my PR to add the context-1m-2025-08-07 beta header for non-OAuth.

P.S. Worth noting that the 200k limit on Opus 4.6 without the beta header is different from the 200k context window on models like Haiku 4.5. For Haiku 4.5, 200k is the total context window — input tokens, output tokens, everything has to fit within that 200k budget. For Opus 4.6, the context window is actually 1M — the 200k is only a gate on input tokens. Output and thinking tokens live in the larger 1M space and don't count against the 200k. So even without the context-1m-2025-08-07 header, Opus 4.6 gives you significantly more effective room. For example, if you're using extended thinking with large thinking budgets, those tokens aren't eating into your 200k input limit like they would eat into Haiku 4.5's 200k context window. And thinking tokens from previous turns are automatically stripped by the API, so they don't accumulate as input at all.

[okhsunrog]

@rekram1-node @thdxr could anyone review this, please?

[BouquetAntoine]

Work well for me except the displayed usage is wrong (that's also why you got the 200k limit - claude code limit - warning at "400k" context)

Bug: Token Counter Displays ~2× Actual Usage with @ai-sdk/anthropic v3.x

Symptom

When using @ai-sdk/anthropic v3.x, the token counter displays approximately double the actual token usage (e.g., 49,000 shown vs 24,450 in provider logs).

Root Cause

The SDK v3 changed the structure of LanguageModelUsage.inputTokens:

Version	inputTokens value
v2.x	input_tokens (excludes cache)
v3.x	total = noCache + cacheRead + cacheWrite

In session/index.ts, the getUsage() function assumes that for Anthropic, inputTokens excludes cached tokens:

const excludesCachedTokens = !!(input.metadata?.["anthropic"] || input.metadata?.["bedrock"])

const adjustedInputTokens = excludesCachedTokens
  ? (input.usage.inputTokens ?? 0) // ❌ This is now the TOTAL in v3!
  : ...

Then in the UI (header.tsx, sidebar.tsx), tokens are displayed as:

tokens.input + tokens.output + tokens.reasoning + tokens.cache.read + tokens.cache.write

Result: Cache tokens are counted twice — once inside tokens.input (which is now the total) and again as tokens.cache.read + tokens.cache.write.

Fix

Use the new inputTokenDetails.noCacheTokens field introduced in SDK v3:

// SDK v3: inputTokens is now TOTAL, use noCacheTokens for pure input
const noCacheInputTokens = input.usage.inputTokenDetails?.noCacheTokens

const adjustedInputTokens = noCacheInputTokens !== undefined
  ? noCacheInputTokens
  : (input.usage.inputTokens ?? 0) - cacheReadInputTokens - cacheWriteInputTokens

Also update cache token extraction to use the new structure:

const cacheReadInputTokens =
  input.usage.cachedInputTokens ??
  input.usage.inputTokenDetails?.cacheReadTokens ??
  0

const cacheWriteInputTokens =
  input.usage.inputTokenDetails?.cacheWriteTokens ??
  0

SDK v3 Type Reference

type LanguageModelUsage = {
  inputTokens: number | undefined;          // Now the TOTAL
  inputTokenDetails: {
    noCacheTokens: number | undefined;      // Pure input without cache
    cacheReadTokens: number | undefined;    // Tokens read from cache
    cacheWriteTokens: number | undefined;   // Tokens written to cache
  };
  cachedInputTokens?: number | undefined;   // @deprecated — use inputTokenDetails.cacheReadTokens
  // ...
}

[BouquetAntoine]

can confirm I managed to use the 1M context

[reynard93]

how did anyone manage to use with this branch? i keep getting : The long context beta is not yet available for this subscription

[loop-uh]

+1

[avarayr]

opencode core shouldn't have builtin plugins that pull from forks.

[rekram1-node]

correct

[okhsunrog]

It was only supposed for testing till changes to opencode-anthropoc-auth are accepted