Claude Context Window Size (2026): 1M Tokens on Opus 5.5, Opus 5, Opus 4.8, and Sonnet 5

Every Claude Model's Context Window (Opus 5.5, Opus 5, Sonnet 5, Fable 5.1, Haiku 4.5)

Anthropic's current lineup is Fable 5.1, Opus 5.5, Sonnet 5, and Haiku 4.5. Three of the four have 1M tokens. Opus 5.5 is the newest: it launched September 22, 2026, and replaced Opus 5 (July 24, 2026) as the default Opus. Fable 5.1 launched September 1, 2026. Everything from Claude 4.6 on, except Haiku, carries a 1M window with a 128K synchronous output cap. Current API IDs are claude-opus-5-5, claude-fable-5-1, claude-sonnet-5, and claude-haiku-4-5. Anthropic's Opus 5.5 post says Sonnet 5.5 and Haiku 5.5 will follow in the coming weeks.

For every 1M model, 1M is the default. You do not send a beta header, and a 900K-token request bills at the same per-token rate as a 9K one. On the Message Batches API, Opus 5.5, Opus 5, Opus 4.8, Opus 4.7, Opus 4.6, Sonnet 5, and Sonnet 4.6 can return up to 300K output tokens with the output-300k-2026-03-24 beta header. Sonnet 5's $2/$10 launched as introductory pricing and is now the standard price; the scheduled September 1 move to $3/$15 was cancelled. Claude Opus 4.1 retired from the Claude API on August 5, 2026, and Opus 4 and Sonnet 4 retired on June 15. Sonnet 4.5 has no formal deprecation notice yet, but its earliest retirement date is September 29, 2026. Sources: Anthropic models overview, pricing, and context windows docs, September 22, 2026.

Two smaller limits sit beside the token limit. A single request can carry up to 600 images or PDF pages on a 1M model, but only 100 on a 200K model. And the Models API returns max_input_tokens and max_tokens for every model, so you can read these limits programmatically instead of hardcoding them.

Claude.ai and Claude Code Context Window Size

The API window is not the window you get in the Claude apps. Anthropic's support article on paid-plan context windows lists three separate tables: claude.ai chat, Claude Code, and Claude Cowork. The chat app halves the window for every model older than the current generation.

In Claude Code, Opus 4.6 and Sonnet 4.6 only get 1M if you pick the [1m] model with /model (for example claude-sonnet-4-6[1m]) and have usage credits enabled (Opus 4.6 needs them on Pro; Sonnet 4.6 on every plan except usage-based Enterprise). Opus 5 had a similar gotcha at launch. On July 25, 2026, Wade Tregaskis reported that Opus 5 defaulted to a 0.2M window on paid plans and that enabling usage credits restored 1M, with no spend required. That report is where the "Opus 5 context window 200K" searches come from. Anthropic's current table lists Opus 5 at 1M on every surface.

A Claude Code session is never empty. Before you type, the system prompt, your CLAUDE.md files, auto memory (the first 200 lines or 25KB of MEMORY.md), skill descriptions, and MCP tool names are already loaded, according to Anthropic's Claude Code context-window guide. After /compact, Claude Code re-reads up to five recently modified files and re-injects each invoked skill, capped at 5,000 tokens per skill. The skill listing itself does not reload. Anthropic's session-management post gives the simplest rule: start a new session for a new task.

What Happens When You Hit the Claude Context Window Limit

Everything counts toward the window: system prompt, every message, tool results, images, documents, tool definitions, and the output Claude generates on the turn, including thinking. Cached tokens count too. Caching changes what you pay for them, not whether they occupy space.

If the input alone exceeds the window, the API returns a 400 invalid_request_error ("prompt is too long") on every model. If input plus max_tokens exceeds the window on Claude 4.5 and newer models, the API accepts the request. Generation stops with stop_reason: "model_context_window_exceeded" if it reaches the limit. Older models return a validation error instead, unless you send the model-context-window-exceeded-2025-08-26 beta header.

For conversations that keep growing, Anthropic offers server-side compaction in beta for Claude 4.6 and later models. It summarizes earlier turns on the server so the conversation continues past the limit. Since September 14, 2026, a separate beta (compact-2026-09-04) lets you trigger compaction on demand with a top-level compaction parameter, which returns a signed compaction block. Context editing can also clear old tool results or thinking blocks. In claude.ai, paid plans with code execution enabled get automatic summarization of earlier messages near the limit.

Context Window vs Effective Context

A context window is how many tokens you can send. Effective context is how many of those tokens the model uses reliably. Anthropic's own docs say it directly: as token count grows, accuracy and recall degrade, a phenomenon known as context rot.

Chroma published one of the most cited studies of this gap on July 14, 2025, testing 18 models including Claude 4, GPT-4.1, Gemini 2.5, and Qwen3. Performance dropped as input length grew on every model, even on simple tasks.

Adobe Research's NoLiMa benchmark (arXiv 2502.05167) sharpens the point. It extends needle-in-a-haystack with needles that share almost no words with the question, so the model has to infer the link instead of string-matching. Of 13 models that claim at least 128K context, 11 dropped below 50 percent of their short-context baseline at 32K tokens. GPT-4o fell from 99.3 percent to 69.7 percent.

The Lost-in-the-Middle Problem

LLMs attend most strongly to tokens at the beginning and end of the context. Information in the middle gets less weight. Put the instructions and the question at the edges of a long prompt, and put the bulk material between them.

Chroma's study added a counterintuitive result: models performed better on shuffled haystacks than on logically structured ones. Coherent documents made retrieval harder, not easier.

NVIDIA's RULER benchmark (arXiv 2404.06654) tested 17 long-context models across 13 tasks. All claimed at least 32K tokens; only half held satisfactory performance at 32K. Anthropic's own long-context number is from the Opus 4.6 launch: 76% on MRCR v2 with 8 needles across 1M tokens, against 18.5% for Sonnet 4.5. Anthropic called it "a qualitative shift in how much context a model can actually use while maintaining peak performance." The Opus 5.5 announcement published no new long-context benchmark.

How Claude Compares to GPT-4.1, Gemini, and Others

Window size no longer separates the frontier. Every current Claude model except Haiku 4.5 sits at 1M, and open-weight models match it: Morph serves GLM-5.3 at 1M and DeepSeek V4 Flash at 1M context. GPT-4.1 is several generations old. OpenAI's current API models are the GPT-6 family.

GPT-6 Astra launched September 3, 2026, and GPT-6 Sol and Luna followed on September 22. Gemini 3.8 Flash went GA on September 2. Grok 4.7 shipped September 21. DeepSeek V4.1 Flash shipped September 10 and replaced V4 Flash on DeepSeek's own API. Kimi K3 reached Moonshot's API in July. OpenAI, Gemini 3.1 Pro, and xAI charge more once a prompt passes a threshold. Claude bills its full 1M window at one rate. DeepSeek V4.1 Flash has the largest output cap at 384K. Grok 4.7 has the smallest window at 500K. Sources: OpenAI pricing, Google's Gemini 3.8 Flash and Gemini 3.1 Pro model pages, xAI models, DeepSeek, and Kimi pricing, September 22, 2026.

Size is close to a tie. The comparison that matters is behavior at length, and the best independent data is still the 2025 studies.

Chroma's model-level results split by family. "Claude models consistently exhibit the lowest hallucination rates" and lean conservative under ambiguity, while "GPT models show the highest rates of hallucination." On the repeated-words task, GPT-4.1 refused 2.55% of attempts and Claude Opus 4 refused about 2.89%. Gemini 2.5 Pro was the most erratic, generating random text starting around 500 to 750 words.

Vendor benchmarks tend to favor the vendor. Needle-in-a-haystack scores measure retrieval of one planted fact, not reasoning across facts spread through the document. Run your own eval at the lengths you actually use.

Prompt Caching: 90% Cost Reduction (97.5% on Fable 5.1)

Prompt caching reuses a processed prompt prefix across API calls. The standard cache read costs 10% of base input. Two new models go lower: Opus 5.5 reads cache at 5% of base input ($0.20/MTok), and Fable 5.1 and Mythos 5.1 at 2.5% ($0.25/MTok). On a 1M-context agent that resends the same repository every turn, cache reads are most of the input bill, so the multiplier matters more than the headline rate.

Anthropic's launch example: a 100K-token book prompt dropped from 11.5 seconds to 2.4 seconds with caching, at 90% lower cost. For a worked number on Sonnet 5, take 100 calls in an hour sharing a 50K-token prefix with the cache kept warm. Uncached, that prefix costs $10.00 per hour. Cached, it costs one $0.125 write plus 99 reads at $0.01 each, about $1.12. Opus 5.5 and Sonnet 5 now charge the same $0.20/MTok for cache reads, so on cache-heavy agents the gap between them is mostly output price. Caching multipliers stack with the 50% Batch discount and the 1.1x US-only inference_geo multiplier.

Extended Thinking and Context (Adaptive Thinking on Opus 5.5 and Sonnet 5)

Thinking tokens are part of max_tokens, bill as output, and count toward the window on the turn they are generated. What happens to them on the next turn depends on the model, and this changed with recent Claude generations.

On Opus 4.5 and later Opus models, Sonnet 4.6 and later Sonnet models, and the Fable and Mythos 5 models, the API keeps previous thinking blocks by default. They count toward the window like any other input and bill as input on later requests. On earlier Opus and Sonnet models and all Haiku models, the API strips previous thinking blocks automatically. If you were counting on stripped thinking to keep a long Opus conversation small, that assumption no longer holds. Use thinking block clearing through context editing to override the default.

Thinking cannot be turned off on Opus 5.5. Sending thinking: disabled, or enabled with budget_tokens, returns a 400. Opus 5 still allowed disabled at effort high or below. Opus 5.5 defaults to effort medium, while Fable 5.1, Sonnet 5, and Opus 5 default to high. Higher effort means more thinking tokens per turn. All 1M models support 128K output per request, and up to 300K on the Batch API with the output-300k-2026-03-24 header on the Opus 4.6+ and Sonnet 4.6+ models listed above.

How Much Code Fits in 200K Tokens (and 1M)

Anthropic's rule of thumb is about 4 characters or 0.75 words per token for English. Code is denser in symbols and whitespace, so treat these as estimates. Assuming about 40 characters per line, a line costs about 10 tokens on the older tokenizer and about 13 on the Claude 4.7+ tokenizer, which produces roughly 30% more tokens.

A React project with 200 files averaging 75 lines (15,000 lines) fits in 200K. A 100,000-line codebase does not fit in 200K. On Opus 5.5 it runs to about 1.3M tokens at 13 tokens per line, so it does not fit in 1M either, before you add the system prompt, tool definitions, and output. Measure with the token counting API before you design around a number.

Cognition (makers of Devin) measured that their coding agent spent 60% of its time searching for code before making changes. The bottleneck was finding the right code to put in context, not the size of the window.

Context Engineering > Context Size

Windows went from 4K to 200K to 1M. Chroma's results show performance falling with length on all 18 models they tested. Bigger windows give you capacity, not quality.

Context engineering is choosing what goes into the window: the relevant code, the right instructions, and nothing that competes with them for attention. Anthropic's context-window docs make the same point: "more context isn't automatically better," and curating what is in context matters as much as how much space is available.

Practical strategies

A 200K window filled with the right code beats a 1M window filled with a whole repository. The tooling that picks and shrinks context decides outcomes more than the window does. Morph Compact does the shrinking. If you want a raw long window on an open-weight model, Morph serves GLM-5.3 (1M) and DeepSeek V4 Flash (1M) through one OpenAI-compatible API.

Frequently Asked Questions

Related Reading

• Context Rot: Why LLMs Get Worse with More Input
• LLM Context Window Comparison (2026)
• LLM Context Windows Explained
• Context Engineering for AI Agents
• Anthropic API Pricing
• Claude Code Token Usage
• Claude Code Usage Limits
• Morph Compact: Context Compression at 33,000 tok/s
• GLM-5.3: 1M-Token Open-Weight Model on Morph
• DeepSeek V4 Flash: 1M Context on Morph
• WarpGrep: Semantic Codebase Search