Claude Opus 4.6/4.7 and Sonnet 4.6 context window understated 5x (200K vs 1M) in MODEL_METADATA; Sonnet 4.6 output cap halved

[sushantlokhande14]

What happened?

MODEL_METADATA in autogpt_platform/backend/backend/blocks/llm.py understates the context window for three current Anthropic models by 5x, and halves one output cap:

Enum	Current metadata	Actual (Anthropic docs)
CLAUDE_4_6_OPUS (claude-opus-4-6)	context 200000	context 1,000,000
CLAUDE_4_7_OPUS (claude-opus-4-7)	context 200000	context 1,000,000
CLAUDE_4_6_SONNET (claude-sonnet-4-6)	context 200000, max output 64000	context 1,000,000, max output 128,000

Sources:

• The comment above the table itself cites https://docs.anthropic.com/en/docs/about-claude/models — which lists Opus 4.6/4.7 and Sonnet 4.6 at a 1M context window and 128K max output.
• The repo's own vendored backend/copilot/anthropic_rates.json (snapshot of litellm's catalog) carries max_input_tokens: 1000000 / max_output_tokens: 128000 for these same models — llm.py contradicts its sibling data file.
• Not a deliberate platform cap: the same table records GPT-4.1 at 1047576 and Grok 4 Fast at 2000000.

Impact

These values are load-bearing in llm.py:

• context_window drives prompt-window budgeting and compression (target_tokens=llm_model.context_window // 2), so conversations on these models get compressed/trimmed at 1/5th of real capacity.
• available_tokens = max(context_window - estimated_input_tokens, 0) strangles the output budget the same way.
• input_data.max_tokens = llm_model.max_output_tokens caps Sonnet 4.6 completions at half their real ceiling.

Steps to reproduce

Select claude-opus-4-6 / claude-opus-4-7 / claude-sonnet-4-6 in an AI block with a long prompt (>200K tokens): the platform compresses/trims it even though the model would accept it as-is.

PR with the three-line metadata fix incoming.

[albatrossflyon-coder]

Confirmed and fixed the context-window part of this (200K -> 1M for all three, verified against Anthropic's docs and this repo's own vendored anthropic_rates.json). One correction though: I checked anthropic_rates.json directly before touching anything, and it actually lists claude-sonnet-4-6's max_output_tokens as 64000, not 128000 — matching what's currently in llm.py, not the number claimed in this issue. So I left max_output_tokens unchanged for all three models (they were already correct). PR branch is ready on my fork but PR creation is currently failing on my end with a GitHub permissions error unrelated to this repo's own settings — will open it as soon as that clears, or happy to have someone else push the fix through if that's faster.