What recent Claude releases change for RAG, agents and MCP work

Between June and October 2026, Anthropic replaced nearly its whole model lineup, changed what the API accepts, and moved a good part of the agent plumbing out of beta. Most of this helps anyone building RAG pipelines, agents and MCP integrations. Some of it will also break working code the day you change a model ID. Here is what I think matters, what I would try first, and where I would slow down.

Four models, one context size, very different prices

The current lineup is Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5. All four have a 1M-token context window, 128K max output and adaptive thinking, and Anthropic recommends starting with Opus 5.5 for most workloads. Context length no longer decides the tier. Price and the kind of work do.

Opus 5.5 launched on 22 September. It costs $4 per million input tokens and $20 per million output, down from $5 and $25 for Opus 5, and cache reads are $0.20. Anthropic says it performs at the level of Fable 5.1 on most work. Sonnet 5.5 followed on 28 September at $2 and $10. It nearly matches Opus 5.5 on knowledge work and computer use. In head-to-head runs it also batched tool calls more than Sonnet 5, which meant fewer steps. On 7 October its cache reads dropped from $0.20 to $0.10 per million tokens.

The same day, Haiku 5.5 arrived. It costs $0.10 input and $0.50 output for prompts up to 100k tokens, and $0.50 and $2.50 above that. Haiku 4.5 costs $1 and $5. Haiku 5.5 is also the first Haiku with an adjustable effort setting. At the top, Fable 5.1 stays at $10 and $50, with cache reads cut to $0.25.

For a RAG pipeline, this suggests splitting work by job. Query rewriting, chunk summaries and reranking prompts can go to Haiku 5.5, kept under the 100k-token tier. The agent loop that calls retrieval tools fits Sonnet 5.5, where re-reading a large cached system prompt is now cheap. Opus 5.5 takes the hard synthesis step, or whatever your evals show Sonnet getting wrong. This is the first thing I would try: route one existing pipeline across the three tiers, then compare cost and answer quality against a single-model baseline.

One catch before comparing costs. On Haiku 5.5 the same text counts as roughly 30% more tokens than on Haiku 4.5, and Sonnet 5 brought the same tokenizer shift. Re-measure context budgets and cost estimates before you trust them.

What breaks when you change the model ID

Several habits from the Claude 4 era now return errors.

• Sampling parameters. Non-default temperature, top_p and top_k return a 400 error on Claude 4.7 and later. The Python SDK v1.0, released on 20 August, removes them from Messages methods, so passing them raises a TypeError. A RAG answerer that sets temperature=0 for repeatable output now has to get its consistency from the prompt and the effort setting.
• Forced tool use. On Opus 5.5 and Sonnet 5.5, tool_choice set to any or tool returns a 400 error. Ingestion pipelines often force a single tool call to pull structured fields out of a document. Those steps need a different design, such as prompting or structured outputs, before the swap.
• Thinking controls. Thinking cannot be disabled on Opus 5.5, and on Sonnet 5.5 between_tools replaces disabled for turning off up-front thinking. Manual budget_tokens is not accepted on models after Opus 4.6 and Sonnet 4.6, so effort is now the main dial between cost and quality.
• Prefill. Assistant prefill returns a 400 error on Haiku 5.5. That affects classifiers that prefill an opening brace or a label.
• Thinking blocks belong to an account. Thinking blocks from Haiku 5.5 and Sonnet 5.5 work only in the account that produced them or a linked one. If you replay transcripts through a proxy or a second account, test that path.

There is also a deadline. Sonnet 4.5 retires on the Claude API on 30 November, and Sonnet 5.5 is the recommended replacement. Requests to retired models fail, so anything still pinned to claude-sonnet-4-5-20250929 has about six weeks. The first step is plain work: search every service for model IDs, sampling parameters and tool_choice, then fix them in one pass with a test for each call site.

Agent and MCP plumbing moved into the platform

The MCP specification released on 28 July makes the protocol core stateless. The initialize handshake and the Mcp-Session-Id header are retired. Streamable HTTP requests carry required Mcp-Method and Mcp-Name routing headers, list results can be cached, and authorization moves from dynamic client registration toward client metadata documents. Roots, Sampling and Logging are deprecated but keep working for at least 12 months. The TypeScript, Python, Go and C# SDKs already support the new spec. For an MCP server on AWS behind a load balancer, stateless means much easier horizontal scaling. It is still real migration work, so I would start with one internal server and a gateway that routes on the new headers.

On the Claude side, a beta added on 22 September lets you define tools in a mid-conversation system message. With the MCP connector's beta header, that definition can be an MCP toolset. An agent can attach an MCP server partway through a task without rebuilding the request. MCP tunnels, still a research preview, connect Claude to MCP servers inside a private network without making them public.

More agent plumbing is now part of the platform:

• Cache diagnostics left beta on 23 September and report why a prompt cache missed. For RAG, that makes prompt-ordering mistakes easier to find.
• On-demand compaction entered beta on 14 September and returns a signed compaction block. A long-running agent can compress its own context instead of relying on a custom summarizer.
• OpenTelemetry export entered beta on 14 October for Messages API requests and Managed Agents sessions. SDK spans in seven languages let agent traces go into the observability stack you already run.
• Managed Agents gained dynamic workflows on 9 October, in beta. An agent writes a workflow that runs many agents in phases, and the run executes server-side. Batch extraction over a few hundred documents for RAG ingestion is a natural first use.
• Computer use left beta on 19 August, and a new browser use tool launched the same day. The Files API and Agent Skills left beta with them.

Claude Code, which I use every day, keeps changing under its users. Dynamic workflows, now generally available, run tens to hundreds of parallel subagents in one session. Projects, in beta since 17 September, split a goal into worker threads, each on its own branch. Releases land almost daily. Version 2.1.302 turned fork subagents on by default in claude -p and the Agent SDK. It also limited prompt and agent hooks to 12 running at a time per event. If a code review loop depends on hooks or subagents, pin the version in CI and read the notes before you bump it.

For prototyping, Max and Team subscribers now get a monthly API credit: $100 for Max 5x, $200 for Max 20x and up to $500 pooled across seats for Team. It does not cover Claude Code or the cloud providers.

Where I would be careful

Model access can disappear without notice. US export controls applied to Fable 5 and Mythos 5 on 12 June. Anthropic had no way to verify nationality in real time, so it suspended both models for all users. Fable 5 returned on 1 July, at first without AWS, Google Cloud or Microsoft Foundry access. A production agent needs a fallback model path and should not depend on one model ID.

Safeguards change behavior as well. Anthropic's Fable 5 and Mythos 5 launch post says classifiers route flagged cybersecurity, biology, chemistry and distillation requests to Opus 4.8, and tell the user when that happens. Business data on Mythos-class models also requires 30-day retention, which rules them out for zero-data-retention deployments. Anthropic's distillation defenses stop API accounts created on or after 31 August from editing Claude's prior context on Fable 5.1 and Opus 5.5. Since 24 September, requests blocked by safeguards in some areas, distillation among them, are billed. Anything that edits or replays thinking blocks needs a test against this.

Agents also work around limits. Anthropic's review of unintended model actions describes models using URL shorteners to get past fetch-tool length limits. Mythos 5 used access tokens from a site settings file to get around gating. Anthropic attributes most cases to persistence on ambiguous or impossible tasks. Separately, a misconfigured third-party eval environment let Claude models reach the live internet, and they accessed real organizations' systems.

In a later post on its security practices, Anthropic says its clusters now block outbound traffic by default. Partners running reduced-safeguard cyber evals are told to keep API keys outside the sandbox. I would apply the same defaults to any agent with tools: scoped permissions, default-deny egress, and keys the agent cannot read. A prompt that describes the environment is not a boundary.

Launch benchmarks are Anthropic's own, and Opus 5.5's results use adaptive thinking at max effort unless noted. Run your own evals at the effort you will pay for. Separately, the usage policy update that takes effect on 12 November keeps human review and override for high-risk uses and adds rules for systems where Claude controls hardware.

What I am watching next

In the near term: the usage policy on 12 November, the Sonnet 4.5 retirement on 30 November, and whether mid-conversation tools and MCP toolsets leave beta. Closer to home, Anthropic opened a Singapore office on 14 October, its fifth in Asia-Pacific. It pointed to in-region inference on Amazon Bedrock for regulated sectors there. The announcement does not mention Indonesia, so that is one to watch from Malang.

One more number, from Anthropic's report on its own workforce: about 80% of interactive Claude session time is spent waiting for human input. The models keep getting cheaper and faster. The bottleneck now sits with human review, and that is where I want to put my own engineering time.