diff --git a/CLAUDE.md b/CLAUDE.md index efdd518e..6e270569 100644 --- a/CLAUDE.md +++ b/CLAUDE.md @@ -271,6 +271,7 @@ class OutputEvent(BaseEvent): Provider-prefixed strings. Examples: ``` anthropic/claude-fable-5 +anthropic/claude-opus-5 anthropic/claude-sonnet-5 anthropic/claude-opus-4-8 anthropic/claude-opus-4-7 diff --git a/docs/models/anthropic.mdx b/docs/models/anthropic.mdx index 091fceb5..dff46830 100644 --- a/docs/models/anthropic.mdx +++ b/docs/models/anthropic.mdx @@ -1,9 +1,9 @@ --- title: "Anthropic" -description: "Claude Fable 5, Sonnet 5, Opus, and Haiku models with specs, pricing, and capabilities" +description: "Claude Fable 5, Opus 5, Sonnet 5, and Haiku models with specs, pricing, and capabilities" --- -Source: [Anthropic model docs](https://docs.anthropic.com/en/docs/about-claude/models) · [Opus 4.8 pricing](https://docs.anthropic.com/en/about-claude/pricing) +Source: [Anthropic model docs](https://docs.anthropic.com/en/docs/about-claude/models) · [What's new in Opus 5](https://platform.claude.com/docs/en/about-claude/models/whats-new-opus-5) · [Pricing](https://docs.anthropic.com/en/about-claude/pricing) ## Latest @@ -24,6 +24,22 @@ description: "Claude Fable 5, Sonnet 5, Opus, and Haiku models with specs, prici - Knowledge cutoff Jan 2026 + + Reasoning · Speed + + `anthropic/claude-opus-5` + + Latest Opus and the recommended default for complex agentic coding and enterprise work — near-Fable-5 intelligence at half the price, with large gains in deep reasoning, long-horizon tool loops, and test-time compute scaling. + + - $5 / $25 per 1M tokens (input / output; Fast mode $10 / $50 at ~2.5x speed) + - 1M context (default and maximum) + - 128K max output + - Text, Image input + - Adaptive thinking (on by default) + - Web search + - Knowledge cutoff May 2026 + + Reasoning · Speed @@ -45,7 +61,7 @@ description: "Claude Fable 5, Sonnet 5, Opus, and Haiku models with specs, prici `anthropic/claude-opus-4-8` - Latest Opus for agentic coding, long-running tasks, and complex reasoning. Builds on 4.7 with stronger consistency and autonomy across coding, agents, and professional work. + Previous Opus generation for agentic coding, long-running tasks, and complex reasoning. Same price as Opus 5 — migrate unless you depend on thinking being off by default. - $5 / $25 per 1M tokens (input / output) - 1M context diff --git a/docs/models/overview.mdx b/docs/models/overview.mdx index cd5e3742..255417e6 100644 --- a/docs/models/overview.mdx +++ b/docs/models/overview.mdx @@ -7,7 +7,7 @@ Every model supported by Timbal with full specs, pricing, capability scores, and - Claude Fable 5, Sonnet 5, Opus, and Haiku + Claude Fable 5, Opus 5, Sonnet 5, and Haiku Seed 2.0 and Seed 1.8 models diff --git a/python/tests/core/test_frontier_models_integration.py b/python/tests/core/test_frontier_models_integration.py index 8943ecd6..03c5827a 100644 --- a/python/tests/core/test_frontier_models_integration.py +++ b/python/tests/core/test_frontier_models_integration.py @@ -19,6 +19,12 @@ PROMPT = "Reply with exactly one word: ok" LIVE_MODELS = [ + pytest.param( + "anthropic/claude-opus-5", + "ANTHROPIC_API_KEY", + None, + id="anthropic-claude-opus-5", + ), pytest.param( "google/gemini-3.6-flash", "GEMINI_API_KEY", @@ -76,8 +82,9 @@ async def test_frontier_model_agent_collect(model: str, env_key: str, fallback_e _skip_if_no_key(env_key, fallback_env) # Reasoning models (MiniMax / Qwen) spend many tokens on reasoning_content - # before visible text — 64 is often exhausted with an empty content array. - agent = Agent(name=f"probe_{model.replace('/', '_')}", model=model, max_tokens=512, tools=[]) + # before visible text, and Opus 5 thinks by default — a small budget is + # often exhausted with an empty content array. + agent = Agent(name=f"probe_{model.replace('/', '_')}", model=model, max_tokens=2048, tools=[]) result: OutputEvent = await agent(prompt=PROMPT).collect() assert result.status.code == "success", result.error @@ -99,7 +106,7 @@ async def test_frontier_model_llm_router_streams(model: str, env_key: str, fallb async for chunk in _llm_router( model=model, messages=[Message(role="user", content=[TextContent(text=PROMPT)])], - max_tokens=512, + max_tokens=2048, ): chunks.append(chunk) diff --git a/python/timbal/core/models.py b/python/timbal/core/models.py index adfde550..28b40ccf 100644 --- a/python/timbal/core/models.py +++ b/python/timbal/core/models.py @@ -44,6 +44,7 @@ def get_context_window(model_id: str) -> int | None: # Model type with provider prefixes Model = Literal[ "anthropic/claude-fable-5", + "anthropic/claude-opus-5", "anthropic/claude-opus-4-8", "anthropic/claude-sonnet-5", "anthropic/claude-opus-4-7", diff --git a/python/timbal/models.yaml b/python/timbal/models.yaml index de283019..b16a7467 100644 --- a/python/timbal/models.yaml +++ b/python/timbal/models.yaml @@ -14,16 +14,27 @@ models: display_name: Claude Fable 5 description: Anthropic's most capable widely released model for long-horizon agentic work, complex reasoning, and ambitious coding tasks. - notes: Includes safety classifiers; blocked requests return stop_reason refusal. Configure Opus 4.8 fallback for production + notes: Includes safety classifiers; blocked requests return stop_reason refusal. Configure Opus 5 fallback for production workloads. input_price: 10.0 output_price: 50.0 context_window: 1000000 capabilities: [vision, tools, reasoning] +- id: anthropic/claude-opus-5 + provider: anthropic + display_name: Claude Opus 5 + description: Anthropic's latest Opus — near-frontier intelligence at half the price of Fable 5, with step-change gains in + deep reasoning, agentic coding, and long-horizon tasks. + notes: Thinking is on by default; effort defaults to high and `thinking.type=disabled` is rejected at xhigh/max effort. + Fast mode (~2.5x speed) is $10/$50 per 1M tokens on the Claude API only. + input_price: 5.0 + output_price: 25.0 + context_window: 1000000 + capabilities: [vision, tools, reasoning] - id: anthropic/claude-opus-4-8 provider: anthropic display_name: Claude Opus 4.8 - description: Anthropic's latest Opus — improved coding, agents, and long-context reasoning. + description: Previous-generation Opus for coding, agents, and long-context reasoning. Superseded by Opus 5 at the same price. input_price: 5.0 output_price: 25.0 context_window: 1000000 @@ -67,7 +78,7 @@ models: provider: anthropic display_name: Claude Opus 4.1 description: Previous-generation Opus model with high intelligence for complex reasoning tasks. - notes: Deprecated; retires August 5, 2026. Prefer claude-opus-4-8 or claude-fable-5. + notes: Deprecated; retires August 5, 2026. Prefer claude-opus-5 or claude-fable-5. input_price: 15.0 output_price: 75.0 context_window: 200000