Claude / Anthropic
The agent-era platform — best-in-class harness, context engineering and MCP — but you bring your own embeddings, vector store and fine-tuning.
Element scores
Strengths
The strongest agent stack in the industry: Claude Code, the Agent SDK and hosted Managed Agents share one battle-tested harness with hooks, subagents, skills, credential vaults and session-level tracing. Context engineering (1M window, caching, compaction, memory) is best-in-class, Fable 5 is a Mythos-class frontier model, and Anthropic authored MCP — the interoperability standard everyone else now adopts.
Honest dings
The platform is deliberately narrow: no first-party embeddings (docs send you to Voyage AI), no hosted vector store, fine-tuning only via Haiku SFT on Bedrock, and no image or audio generation. Fable 5's safety classifiers can refuse requests, so integrations need explicit refusal-and-fallback handling. Consumer Max tiers are expensive, and the pace of new surfaces (Managed Agents is still beta) means building on moving ground.
Sources
Every audit lists the research it rests on — transparency and traceability are the product. Tools evolve: each audit is a snapshot of its audit date, and re-audits supersede older versions (kept below for reference).
- platform.claude.com/docs/en/about-claude/models/int… — Official: Fable 5 GA 09/06/2026, Mythos-class capabilities, safety classifiers, refusal/fallback/billing changes, context window (accessed 2026-07-07)
- claude.com/pricing — Official pricing page: Pro $20 ($17 annual), Max from $100, Team $20-25/$100-125 per seat, Enterprise seat+usage (accessed 2026-07-07)
- support.claude.com/en/articles/11049762-choose-a-cl… — Official plan comparison (updated 19/05/2026): Free/Pro/Max 5x/20x/Team/Enterprise tiers and capacity multipliers (accessed 2026-07-07)
- code.claude.com/docs/en/agent-sdk/overview — Agent SDK: built-in tools, hooks, subagents, MCP, permissions, session forking — the Claude Code harness as a library (accessed 2026-07-07)
- claude.com/blog/building-with-claude-managed-agents — Managed Agents (10/06/2026): hosted harness, vaults, session event logs, Console timeline, memory/dreaming, multiagent orchestration, outcomes rubrics (accessed 2026-07-07)
- platform.claude.com/docs/en/managed-agents/mcp-conn… — MCP in Managed Agents: up to 20 servers per agent, vault-based auth, permission policies (beta header managed-agents-2026-04-01) (accessed 2026-07-07)
- platform.claude.com/docs/en/build-with-claude/embed… — Official: 'Anthropic does not offer its own embedding model' — docs redirect to Voyage AI (accessed 2026-07-07)
- platform.claude.com/docs/en/test-and-evaluate/eval-… — Console Evaluation tool: test-case generation, CSV import, side-by-side comparison, 5-point grading, prompt versioning (accessed 2026-07-07)
- platform.claude.com/docs/en/build-with-claude/vision — Vision capabilities and limits; official confirmation Claude cannot generate or edit images (accessed 2026-07-07)
- platform.claude.com/docs/en/agents-and-tools/tool-u… — Tool use: client/server tools, strict schemas, parallel calls, programmatic tool calling, tool search (accessed 2026-07-07)
- venturebeat.com/technology/anthropic-says-it-hit-a-… — Independent (08/05/2026): $30B run-rate, Claude Code >$2.5B run-rate, >1,000 enterprise customers ≥$1M/yr, 32% enterprise API share claim (accessed 2026-07-07)
- gate.com/news/detail/anthropic-raises-65b-at-965b-v… — Independent report of official statements (26/06/2026): $47B run-rate, $65B Series H at $965B valuation (accessed 2026-07-07)
- callsphere.ai/blog/vw8g-anthropic-claude-fine-tunin… — Independent (03/2026): no fine-tuning via Anthropic's public API; only supported path is Haiku SFT on Amazon Bedrock (accessed 2026-07-07)
- techcrunch.com/2026/06/09/anthropics-claude-fable-5… — Independent coverage of Fable 5 launch and its guardrail classifiers (accessed 2026-07-07)