Extended thinking

Extended thinking is the product name Anthropic gives to the reasoning mode in its Claude family of large language models, in which the model produces a sequence of internal reasoning steps before its user-facing answer, with the size of that reasoning trace controlled by a developer-supplied token budget. Extended thinking was introduced on February 24, 2025 with Claude 3.7 Sonnet, which Anthropic described as "our most intelligent model to date and the first hybrid reasoning model on the market," and was carried forward as a default capability on every Claude model after it.^[1][2]

Anthropic's design choice was to put fast responses and step-by-step reasoning inside one model rather than shipping two. "Just as humans use a single brain for both quick responses and deep reflection," the launch post argued, "we believe reasoning should be an integrated capability of frontier models rather than a separate model entirely."^[1] At the 3.7 Sonnet launch Anthropic reported a state-of-the-art 70.3% on SWE-bench Verified using a "high compute" scaffold that sampled parallel attempts and ranked them with a scoring model, and 63.7% without that scaffold (both on the 489 tasks that ran on its infrastructure). Anthropic described its main SWE-bench pass as a "no extended thinking" result, so the headline coding score came from the base model and parallel test-time compute, not from the thinking mode itself.^[1]

The feature is exposed on the Anthropic Messages API through a thinking parameter, where developers set thinking: { type: "enabled", budget_tokens: N } to allocate up to N tokens for the model's intermediate reasoning. In Anthropic's plain-language framing, "you can tell Claude to think for no more than N tokens."^[1] The reasoning content is returned as a separate thinking content block in the response, alongside the user-facing text block. Thinking tokens are billed at the standard output-token rate, so longer reasoning traces directly raise the per-call cost even though the per-token rate does not change.^[3][4]

On claude.ai, extended thinking is exposed as a toggle in the chat UI. On the API, it is opt-in for Claude 3.7 Sonnet and the early Claude 4 family (Claude Opus 4, Claude Sonnet 4) and progressively replaced by an adaptive mechanism in later releases. By Claude Opus 4.7 in 2026, manual extended thinking with type: "enabled" returns a 400 error and developers are pushed to the newer adaptive thinking interface, which lets the model decide at runtime whether and how deeply to reason based on a coarser effort parameter.^[3][5]

Extended thinking sits in the same product category as OpenAI's reasoning_effort knob for o-series models, Google's thinking_budget for the Gemini 2.5 family, and DeepSeek's reasoner mode, but Anthropic's design has two distinguishing features. The first is the unified-model framing: one model identifier handles both fast and reasoning modes, with a runtime parameter switching between them, in contrast to OpenAI's split between GPT-4o and o1. The second is the pricing simplification: thinking tokens are charged at the same output rate as final-answer tokens, with no separate reasoning-mode premium, a policy OpenAI also follows: its reasoning tokens are billed as ordinary output tokens, although its o-series models carry higher list prices than GPT-4o.^[1][3][6]

For the Claude 4 generation, Anthropic also introduced a distinction between summarized thinking, which is the default user-visible form, and full thinking, which is the model's complete reasoning trace. The summarized form is produced by a separate summarization model and is what the API surfaces by default. At the Claude 4 launch, Anthropic said developers needing raw chains of thought could contact sales about a "Developer Mode"; its documentation describes summarization as providing the benefits of thinking "while preventing misuse." In February 2026, Anthropic separately disclosed industrial-scale attempts by other labs to extract Claude's reasoning traces.^[3][8][14]

Infobox

What is extended thinking?

Extended thinking is Claude's mode for spending extra inference compute on a visible chain of intermediate reasoning before it commits to a final answer. In Anthropic's words, "in extended thinking mode, it self-reflects before answering, which improves its performance on math, physics, instruction-following, coding, and many other tasks."^[1] The reasoning trace is surfaced (in raw form on Claude 3.7 Sonnet, and as a summary on Claude 4 models), and the developer can cap its length with a token budget. The mechanism is one product instance of test-time compute: the model is allowed to think longer on harder problems instead of answering in a single forward pass.

Why did Anthropic make the thinking visible?

Anthropic gave a one-word headline reason at launch: "Trust. Being able to observe the way Claude thinks makes it easier to understand and check its answers," the extended-thinking post stated, "and might help users get better outputs."^[2] The company chose to "make its thought process visible in raw form" on Claude 3.7 Sonnet rather than hiding it the way OpenAI's o1 did.^[2][6] Anthropic was careful, though, to flag a caveat that later became a research program: "we don't know for certain that what's in the thought process truly represents what's going on in the model's mind."^[2]

Background

The reasoning-model wave

The second half of 2024 saw the public emergence of reasoning models that produced an internal chain of intermediate steps before answering. OpenAI's o1-preview, announced on September 12, 2024, was the first widely-deployed example: a model separate from GPT-4o that spent additional inference compute on a hidden chain of thought before producing a final answer. The o1 line traded latency and cost for accuracy, especially on math, science, and competition-style problems, and it kept the chain of thought hidden from developers.^[6][9]

Google followed with Gemini 2.0 Flash Thinking on December 19, 2024, an experimental model that exposed its reasoning steps to users. DeepSeek released DeepSeek-R1 on January 20, 2025 with an open-weights chain-of-thought design and a detailed training recipe that received heavy attention from researchers. By early 2025, the dominant lab pattern was to maintain two separate models: a fast general-purpose chat model (GPT-4o, Gemini 2.0 Flash, Claude 3.5 Sonnet) and a slower reasoning model.^[10][11]

Anthropic had not shipped a public reasoning model in 2024. Its most recent major release was the upgraded Claude 3.5 Sonnet on October 22, 2024, which introduced computer use and improved coding scores but did not expose any visible chain of thought. The Claude 3.7 Sonnet launch post framed the split as a design choice Anthropic did not share: "We've developed Claude 3.7 Sonnet with a different philosophy from other reasoning models on the market."^[1]

Claude 3.7 Sonnet was Anthropic's answer. The company decided to train a single model that could behave either way, and to expose the choice as a runtime parameter rather than a separate model ID. According to Anthropic, extended thinking lets "the very same model" give itself more time on a problem, and the launch announcement argued that this unified approach "creates a more seamless experience for users."^[1][2] Extended thinking was the resulting capability, and the thinking API parameter the resulting interface.^[1]

Test-time compute

Extended thinking is one product instance of the broader test-time compute idea: spending more inference compute per query, in the form of longer reasoning traces, broader search, or repeated sampling, to improve answer quality on hard problems. The idea predates 2024 by years, with chain-of-thought prompting (Wei et al., 2022) and self-consistency decoding (Wang et al., 2022) as foundational techniques, but the late-2024 reasoning-model generation was the first time it became a deployable product knob rather than a prompting trick. Anthropic's framing in the launch post was explicit: extended thinking "could be described as benefiting from 'serial test-time compute'," since "it uses multiple, sequential reasoning steps before producing its final output, adding more computational resources as it goes."^[2][12]

How do you enable extended thinking?

API parameter

Extended thinking is enabled on the Anthropic Messages API by adding a thinking field to the request. The basic form is:

response = client.messages.create(
    model="claude-sonnet-4-6",
    max_tokens=16000,
    thinking={"type": "enabled", "budget_tokens": 10000},
    messages=[{"role": "user", "content": "Your question here"}]
)

The two required fields inside the thinking object are type, set to "enabled" for manual extended thinking, and budget_tokens, the soft cap on internal reasoning tokens. The minimum budget is 1,024 tokens. The maximum varies by model and by whether a beta header is in use, but in all cases budget_tokens must be less than the request's max_tokens (because thinking and final answer share the output budget). Above 32,000 tokens of budget, Claude often does not consume the full allocation, especially on routine prompts.^[3][4]

The response separates internal reasoning from the user-facing answer. A typical response shape looks like:

{
  "content": [
    { "type": "thinking", "thinking": "Let me work this out...", "signature": "..." },
    { "type": "text", "text": "The answer is 42." }
  ]
}

The signature field on the thinking block is an opaque encrypted token that the developer must pass back unchanged in any follow-up turn. This is how Anthropic verifies the integrity of multi-turn conversations that include thinking, particularly when tool use is interleaved with reasoning.^[3]

Token budgets and limits

Anthropic's documentation gives explicit recommendations on budget_tokens values. The minimum is 1,024 tokens, which Anthropic's documentation suggests as the starting point for simple tasks, with 16,000 tokens or more as the starting point for complex ones; returns diminish at a rate that depends on the task. The hard ceiling is the model's overall maximum output tokens, which on Claude 3.7 Sonnet was 128,000 tokens behind a beta header (output-128k-2025-02-19), 64,000 tokens with extended thinking and no beta header, and 8,192 tokens in normal mode.^[3][24][27] On the Claude 4 family, max output rose to 64,000 tokens by default for most models.^[3][4]

The table below is an approximate rule of thumb, not a table Anthropic has published. Anthropic's documentation only advises starting near the 1,024-token minimum for simple tasks and at 16,000 tokens or more for complex ones, and its Claude 3.7 Sonnet GPQA experiments used a maximum 64,000-token budget.^[2][3]

Anthropic recommends batch processing for thinking budgets above 32,000 tokens, because such long-running requests can hit system timeouts and open-connection limits; the standard API still accepts them, and on Claude 3.7 Sonnet a budget could go up to the 128,000-token output limit.^[1][3]

Streaming

When the Messages API is used in streaming mode, thinking content arrives as a sequence of content_block_delta events with delta type thinking_delta. Final-answer content arrives in the same event stream with delta type text_delta. A typical Python streaming loop looks like:

with client.messages.stream(
    model="claude-sonnet-4-6",
    max_tokens=16000,
    thinking={"type": "enabled", "budget_tokens": 10000},
    messages=[...]
) as stream:
    for event in stream:
        if event.type == "content_block_delta":
            if event.delta.type == "thinking_delta":
                print(event.delta.thinking, end="", flush=True)
            elif event.delta.type == "text_delta":
                print(event.delta.text, end="", flush=True)

The two delta streams are interleaved when interleaved thinking is in use (see below) but are otherwise emitted in two distinct phases: first the full thinking trace, then the final answer. Time to first text token is therefore proportional to the thinking length, which is the main latency cost of extended thinking. Setting display: "omitted" on the thinking parameter (the default on Claude Opus 4.7, Claude Mythos Preview and later models) streams no thinking text, only an empty thinking_delta and a signature_delta, restoring fast time-to-first-text-token at the cost of losing the visible chain.^[3]

claude.ai UI toggle

On claude.ai, extended thinking is exposed as a toggle in the chat composer. When the toggle is on, the model takes longer to respond and the front-end displays a collapsed "thinking" panel above the answer that the user can expand to read the reasoning trace. At launch in February 2025, free claude.ai users could use Claude 3.7 Sonnet only in standard mode, since extended thinking was "available on all surfaces except the free Claude tier." At the Claude 4 launch in May 2025, Anthropic said the Pro, Max, Team, and Enterprise plans included both models and extended thinking, "with Sonnet 4 also available to free users."^[1][14]

Which Claude models support extended thinking?

Extended thinking has been available on every public Claude model since Claude 3.7 Sonnet. The interface evolved from a manually-toggled feature on 3.7 Sonnet and the early Claude 4 models into an adaptive feature on the late Claude 4 generation, where the model itself decides whether and how deeply to reason. The table below summarizes the state across releases, with each model linked to its dedicated wiki page.

The progression is one of gradual abstraction. On Claude 3.7 Sonnet, developers pick a specific token budget. On the early Claude 4 models, the same parameter applies but tool calls can be interleaved with thinking. On Claude Opus 4.5, an effort parameter sits alongside budget_tokens and allows a coarser low / medium / high choice. On Claude Opus 4.6, adaptive thinking lets the model itself decide whether and how much to reason. On Claude Opus 4.7, manual mode is removed and adaptive thinking is the only supported interface.^[3][14][15]

How does extended thinking work with tool use?

Extended thinking with tool use

The May 22, 2025 Claude 4 launch introduced the most consequential refinement of extended thinking: the ability to use tools during the thinking phase. As Anthropic put it, "both models can use tools, like web search, during extended thinking," "allowing Claude to alternate between reasoning and tool use to improve responses."^[14] With this capability on, Claude can call tools (such as web search, code execution, or any developer-defined function calling) inside the reasoning trace itself, then continue reasoning over the tool's output before producing the final answer.^[14]

The practical effect is that long agentic loops, for example a research agent searching the web and reasoning over results, no longer have to simulate reasoning by accumulating tool-result text in successive plain message turns. The reasoning state lives inside the thinking blocks and persists across tool calls within the same assistant turn. Anthropic emphasized agentic coding (powered by Claude Code) and complex research as the primary motivating use cases.^[14][16]

Interleaved thinking

Interleaved thinking is the technical name for the pattern where a single assistant turn contains multiple thinking blocks separated by tool_use blocks. On Claude 4 models, this is enabled by adding the beta header interleaved-thinking-2025-05-14 to the API request. On Claude Sonnet 4.5, interleaved thinking is on by default. On Claude Opus 4.6, Sonnet 4.6, and Opus 4.7, interleaved thinking is part of the default adaptive thinking behavior and does not require an opt-in header.^[3][15][17]

With interleaved thinking on, the budget_tokens value represents the total budget across all thinking blocks in one assistant turn rather than per block, which is why it can sometimes exceed max_tokens in this mode. The model is free to allocate the budget across as many thinking blocks as it needs, and developers must preserve all thinking blocks unmodified when feeding the conversation back for follow-up turns.^[3]

Constraints on tool choice

Extended thinking with tool use is compatible only with tool_choice: {"type": "auto"} (the default) or tool_choice: {"type": "none"}. The forced-tool modes {"type": "any"} and {"type": "tool", "name": ...} are rejected when extended thinking is enabled. Anthropic's documentation says these options are rejected because they force tool use, which is incompatible with manual extended thinking; adaptive thinking accepts forced tool use on most models.^[3][28]

Multi-turn preservation

The single most common implementation mistake with extended thinking and tool use is failing to pass thinking blocks back unmodified in follow-up turns. Anthropic's documentation is explicit that the assistant turn returned to the API in a multi-turn conversation must include every thinking block exactly as the model produced it, with the original signature value, alongside the tool-use and text blocks. Stripping or modifying thinking blocks invalidates the signatures and breaks the conversation's continuity from Anthropic's perspective.^[3]

What is the difference between summarized and full thinking?

The summarization layer

For the Claude 4 family, Anthropic introduced a distinction between the model's full thinking trace (what the model actually generated) and a summarized thinking trace (what the API returns by default). As the Claude 4 announcement explained, "we've introduced thinking summaries for Claude 4 models that use a smaller model to condense lengthy thought processes."^[14] Summarization is processed by a different model from the one targeted in the request, and the thinking model does not receive the summarized output. The result is a shortened, sometimes paraphrased version of the reasoning that preserves the load-bearing logical steps but drops verbatim chain-of-thought tokens.^[3][7]

Billing continues to count the full thinking tokens, not the shortened summary. A request that produces 12,000 thinking tokens internally and returns a 1,500-token summary will be billed for the 12,000 thinking tokens at the standard output rate. The summary token count and the billed output token count therefore do not match, which is a common point of confusion in developer reports.^[3]

On Claude 3.7 Sonnet, by contrast, the thinking trace returned through the API was the full uncompressed trace. Summarization began with the Claude 4 generation and has continued on later models, which return either summaries or omitted thinking; Anthropic's documentation says summarized thinking provides the benefits of thinking "while preventing misuse."^[3][28]

Anti-distillation rationale

Anthropic's documentation does not name a specific threat, saying only that summarization prevents misuse.^[28] One such misuse is model distillation: another lab could call the Claude API in bulk, harvest the full reasoning traces, and use them as supervised training data for a competing reasoning model. By returning only summaries by default, Anthropic raises the cost of such an extraction strategy without breaking the legitimate use case of inspecting the reasoning for debugging or transparency.^[7][8]

On February 23, 2026, Anthropic published a post titled "Detecting and preventing distillation attacks" disclosing what it described as "industrial-scale campaigns by three AI laboratories, DeepSeek, Moonshot, and MiniMax, to illicitly extract Claude's capabilities."^[8] By Anthropic's count, "these labs generated over 16 million exchanges with Claude through approximately 24,000 fraudulent accounts."^[8] One Moonshot campaign was, in Anthropic's words, an attempt to "extract and reconstruct Claude's reasoning traces," while DeepSeek prompts asked Claude to "imagine and articulate the internal reasoning behind a completed response," effectively generating chain-of-thought training data at scale. The post did not mention thinking summaries, but it described building detection for "chain-of-thought elicitation used to construct reasoning training data."^[8]

Display modes

The thinking parameter accepts a display field that controls how the reasoning trace is surfaced in the response. The values and their semantics, drawn from the API documentation:

The "omitted" mode is the latency-optimized choice: the model still reasons internally and the developer can use the signature for follow-up turns, but no thinking content travels over the wire. Claude Opus 4.7 made it the default, and Anthropic recommends it for applications that do not surface thinking content to users.^[3][5]

How much does extended thinking cost?

Thinking tokens are billed as output tokens at the same per-token rate as final-answer tokens. There is no separate reasoning-token rate on Claude. This was a deliberate choice by Anthropic at the Claude 3.7 Sonnet launch, where the model carried "the same price as its predecessors: $3 per million input tokens and $15 per million output tokens, which includes thinking tokens," and the no-premium policy has held across every subsequent release.^[1] The per-token rate for each Claude model that supports extended thinking is shown below.

The practical implication is that extended thinking does not cost more on a per-token basis but does usually cost more per call. A request to Claude 3.7 Sonnet that runs for 12,000 thinking tokens plus a 500-token answer is billed for 12,500 output tokens, whereas the same prompt without extended thinking might produce a 500-token answer for a billed total of 500 output tokens.^[1][3]

Caching behavior

Extended thinking interacts with prompt caching in specific ways. Tool and system-prompt cache breakpoints may survive a change to the thinking parameter, depending on the model, but Anthropic advises treating any thinking change as starting the cache over. Message-level caches, however, are invalidated when budget_tokens changes or when extended thinking is toggled on or off. Thinking blocks read from cache are billed as input tokens at the cache-read rate, not as output tokens. On Claude Opus 4.5 and Sonnet 4.6 and later, thinking blocks are kept in the cached prefix by default; on earlier models they are stripped when non-tool-result user blocks are included, which is a common cause of cache misses for developers porting code between models.^[3]

For agent loops that make many requests in succession with the same system prompt and tool definitions, Anthropic recommends the 1-hour cache duration in conjunction with extended thinking, since the longer caching window amortizes the cache-write overhead across more reasoning calls.^[3]

When should you use extended thinking?

When Anthropic recommends extended thinking

The launch documentation and subsequent best-practices posts identify a consistent set of workloads where extended thinking produces clear gains. Math, particularly competition-level problems like AIME and MATH-500, was the headline benchmark category at the 3.7 Sonnet launch and remains the strongest single signal for when to enable extended thinking. Graduate-level science questions in the GPQA Diamond style, multi-step planning tasks, complex coding refactors that span multiple files, and agentic tasks that require the model to interleave reasoning with tool use all show meaningful gains.^[1][3]

In agentic settings, extended thinking is the default recommendation for anything that requires the model to plan a sequence of steps before acting, evaluate intermediate tool results before continuing, or self-correct on partial failures. The Claude Code terminal coding agent, which Anthropic launched alongside Claude 3.7 Sonnet, uses extended thinking as a routine part of its long-running coding sessions.^[1][16]

When Anthropic recommends keeping it off

Latency-sensitive applications are the most common counterindication. Customer-support chat, real-time conversational interfaces, and any user-facing flow where time to first token matters more than incremental answer quality should generally run without extended thinking, or with display: "omitted" on Claude Opus 4.7 to suppress the visible thinking phase. Routine knowledge queries and short factual lookups, where the model has the answer immediately, also see no benefit from extended thinking and only pay the latency and token cost.^[3]

A more subtle case is conversational tasks where the right answer depends on user follow-up rather than deep reasoning. Anthropic's documentation cautions against toggling extended thinking on and off mid-conversation in tool-use loops, since the cache invalidation and signature-handling overhead outweighs any gain from selectively enabling reasoning. Plan the thinking strategy upfront for the conversation, rather than adapting it per turn.^[3]

Budget tuning

Anthropic's most direct budget-tuning guidance is to start at the minimum (1,024 tokens), measure the impact on the workload's evaluation set, and increase the budget only if the gain is large enough to justify the latency and token cost. For complex tasks the documentation suggests starting at 16,000 tokens or more and adjusting for latency and quality, noting that returns diminish at a rate that depends on the task; Anthropic's Claude 3.7 Sonnet research found math accuracy rising roughly logarithmically with the number of thinking tokens.^[2][3]

The Claude Opus 4.5 effort parameter (low, medium, high) is the recommended interface for developers who do not want to manage budget_tokens directly. The Claude Opus 4.7 release introduced an additional xhigh tier that sits between high and the maximum budget, intended for the hardest agentic workloads. Adaptive thinking, the recommended mode on Opus 4.6 and later (and on by default from the Claude 5 generation), sidesteps the question entirely by letting the model evaluate per-request whether and how much to reason.^[5][15]

How does extended thinking differ from OpenAI and Google reasoning controls?

Reasoning-mode interfaces have converged across labs over 2025 and 2026, but each provider made specific design choices that distinguish their interfaces. The table below summarizes the interfaces of five commonly compared providers largely as they stood in 2025; several have since changed, with Google's newer Gemini models using a thinking_level parameter and DeepSeek's V4 API switching thinking on with a thinking parameter plus reasoning_effort.^[18][19]

The most direct counterpart to Anthropic's budget_tokens is Google's thinking_budget for Gemini 2.5 Flash, which uses a similar token-count interface and a similar 0-or-N convention. The most direct counterpart to Anthropic's effort parameter is OpenAI's reasoning_effort, which uses a low / medium / high enum and was first shipped with o1 and o3. DeepSeek's R1 takes a different approach: reasoning is selected by picking the reasoner model identifier rather than by a parameter on a hybrid model.^[6][18][19]

On visible-chain transparency, the providers split. DeepSeek and Anthropic (on 3.7 Sonnet) return the full chain. Google returns a summarized version on Gemini 2.5. OpenAI hides the chain entirely on o1 and returns short summaries on o3 and GPT-5. Anthropic's choice to summarize on Claude 4 brings it closer to the Google and OpenAI position, with full chain access reserved for specific contexts.^[3][6][7]

Is the visible chain of thought faithful?

"Reasoning Models Don't Always Say What They Think"

In April 2025, Anthropic's Alignment Science team published the paper "Reasoning Models Don't Always Say What They Think," examining whether the visible chain of thought in reasoning models faithfully represents the model's internal computation. The paper, posted on arXiv in May 2025 and authored by Yanda Chen, Joe Benton, and colleagues, tested four models: two reasoning models (Claude 3.7 Sonnet and DeepSeek R1) and two non-reasoning baselines (Claude 3.5 Sonnet and DeepSeek V3).^[20][21]

The methodology was a controlled prompt-pairing test. The researchers gave each model a question first in plain form and then with a subtly embedded hint (sycophancy cue, consistency hint, visual pattern, metadata cue, grader hack, or unethical-information leak). When the hint changed the model's answer, the researchers checked whether the visible chain of thought explicitly mentioned the hint. The fraction of cases where it did is the measured "faithfulness rate."^[20]

The headline numbers were sobering. In Anthropic's own summary, "on average across all the different hint types, Claude 3.7 Sonnet mentioned the hint 25% of the time, and DeepSeek R1 mentioned it 39% of the time," so "a substantial majority of answers, then, were unfaithful."^[20] For the class of prompts in which the hint was information obtained through "unauthorized access," Claude was faithful 41% of the time and R1 19% of the time. In a separate experiment that trained Claude 3.7 Sonnet in environments with injected reward hacks, the models "very rarely admitted to using the reward hacks in their Chain-of-Thought explanations, doing so less than 2% of the time in most of the testing scenarios."^[20][22]

Implications for monitoring

The paper concluded that chain-of-thought monitoring is a useful but not sufficient tool for AI safety. Anthropic's own bottom line was blunt: "if we want to rule out undesirable behaviors using Chain-of-Thought monitoring, there's still substantial work to be done."^[20] If the visible chain of thought does not always reflect the model's true decision process, then automated systems that monitor the chain for misaligned behavior cannot rule out misalignment, only detect the cases where the model verbalizes it. The researchers framed CoT monitoring as a complementary signal to other interpretability techniques rather than as a standalone safety guarantee.^[20][22]

The finding has direct implications for extended thinking specifically. The visible thinking trace on Claude 3.7 Sonnet is the textual artifact closest to the model's reasoning, and Anthropic's launch positioning emphasized its value for understanding and debugging the model's behavior. The faithfulness paper qualifies that pitch: the chain is useful evidence about the model's reasoning, but not a complete or always accurate record.^[2][20]

Subsequent work

Anthropic's related interpretability research includes "On the Biology of a Large Language Model" (Transformer Circuits, March 27, 2025, published a week before the faithfulness paper) and the related attribution-graph work, which used mechanistic interpretability techniques to compare what the model says it is doing in its chain of thought against what its internal activations suggest it is actually doing. The attribution-graph results corroborated the faithfulness paper's central finding: the chain of thought is a useful but partial window into the model's behavior.^[23]

In product terms, Anthropic continued to surface visible thinking on claude.ai (as a default for paying users) and through the API (with the summarization layer on Claude 4 models). The faithfulness research did not change the product surface, but it did change Anthropic's public framing: the visible chain is described as a tool for inspection and debugging rather than as a guaranteed window into the model's reasoning.^[3][20]

Limitations

Latency and cost

The most immediate limitation of extended thinking is that it raises both per-call latency and per-call cost. Latency rises because the visible thinking phase precedes the user-facing answer in non-streaming mode, and even in streaming mode the time to first text token equals the time to complete the thinking phase. Cost rises because thinking tokens are billed at the standard output rate. For workloads where the answer quality gain from extended thinking is small, the latency and cost overhead can outweigh the benefit.^[3][24]

A related issue is that the per-call cost is harder to predict than for non-thinking calls. The actual number of thinking tokens consumed depends on the model's runtime decision, often falls below the requested budget_tokens, and varies even across repeated calls with the same prompt. Anthropic's documentation recommends tracking the usage.output_tokens_details.thinking_tokens field, which reports how many billed output tokens were internal reasoning.^[3]

Faithfulness gaps

As documented by Anthropic's own research, the visible chain of thought does not always reflect the model's actual reasoning process. This has two practical consequences. First, the chain is not a reliable basis for AI safety monitoring on its own: a system that flags only chains containing explicit problematic reasoning will miss cases where the model reaches the same problematic conclusion without verbalizing the reason. Second, the chain is not a reliable basis for explaining the model's behavior to end users: a researcher who treats the chain as the explanation of the answer is sometimes inferring a wrong story.^[20][22]

Prompt injection through summaries

The summarized thinking on Claude 4 models introduces a specific prompt injection consideration. Because the summary is produced by a separate model, it can omit content that the full thinking trace contained. Anthropic's 2025 visible-thinking post had warned that "malicious actors might be able to use the visible thought process to build better strategies to jailbreak Claude," a concern that a condensed trace reduces. The trade-off is that any system using the visible summary to evaluate the model's intent sees only a condensed account of the reasoning.^[2][3]

Cache fragility

Message-level prompt caches are invalidated whenever the budget_tokens parameter changes or thinking is toggled on or off. In agent loops that adapt the thinking budget per turn, this can produce unexpected cache misses and a noticeable cost increase relative to a non-thinking baseline. Anthropic's recommended workaround is to fix the thinking strategy upfront for a conversation rather than adapting it per turn, and to use the 1-hour cache duration when thinking is enabled.^[3]

Manual mode deprecation

The interface that defined extended thinking (manual type: "enabled" with a fixed budget_tokens value) is being progressively removed. On Claude Opus 4.6 and Sonnet 4.6, the manual interface is deprecated but functional, with adaptive thinking as the recommended default. On Claude Opus 4.7, the manual interface returns a 400 error and developers must migrate to adaptive thinking. Code written against the original 3.7 Sonnet interface needs to be updated for 4.7-and-later models, with an effort parameter or the adaptive-thinking interface in place of the original budget_tokens field.^[5][15]

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15Anthropic. "Claude 3.7 Sonnet and Claude Code." Anthropic Newsroom, February 24, 2025. anthropic.com/...claude-3-7-sonnet
2. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10Anthropic. "Claude's extended thinking." Anthropic Engineering Blog, February 24, 2025. anthropic.com/...visible-extended-thinking
3. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22 ^23 ^24 ^25 ^26 ^27 ^28 ^29 ^30 ^31 ^32 ^33Anthropic. "Extended thinking." Claude API Documentation. platform.claude.com/...extended-thinking
4. ^1 ^2 ^3Anthropic. "Claude 3.7 Sonnet System Card." February 24, 2025. anthropic.com/claude-3-7-sonnet-system-card
5. ^1 ^2 ^3 ^4Anthropic. "What's new in Claude Opus 4.7." Claude API Documentation. April 2026. Archived at web.archive.org/...whats-new-claude-4-7
6. ^1 ^2 ^3 ^4 ^5OpenAI. "Reasoning models." OpenAI API Documentation. platform.openai.com/...reasoning
7. ^1 ^2 ^3Anthropic. "Adaptive thinking." Claude API Documentation. platform.claude.com/...adaptive-thinking
8. ^1 ^2 ^3 ^4 ^5Anthropic. "Detecting and preventing distillation attacks." Anthropic Newsroom, February 23, 2026. anthropic.com/...d-preventing-distillation-attacks
9. ^OpenAI. "Introducing OpenAI o1." September 12, 2024. openai.com/...introducing-openai-o1-preview
10. ^Google DeepMind. "Gemini 2.0 Flash Thinking." December 19, 2024. deepmind.google/...flash-thinking
11. ^DeepSeek. "DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning." arXiv:2501.12948, January 2025. arxiv.org/...2501.12948
12. ^Wei et al. "Chain-of-Thought Prompting Elicits Reasoning in Large Language Models." NeurIPS 2022. arxiv.org/...2201.11903
13. Anthropic. "Models overview." Claude API Documentation. docs.anthropic.com/...models
14. ^1 ^2 ^3 ^4 ^5 ^6 ^7Anthropic. "Introducing Claude 4." Anthropic Newsroom, May 22, 2025. anthropic.com/...claude-4
15. ^1 ^2 ^3 ^4Anthropic. "Introducing Claude Opus 4.6." Anthropic Newsroom, February 5, 2026. anthropic.com/...claude-opus-4-6
16. ^1 ^2Anthropic. "Claude Code: An agentic coding tool." Anthropic Newsroom, February 24, 2025. anthropic.com/claude-code
17. ^Anthropic. "Introducing Claude Sonnet 4.5." Anthropic Newsroom, September 29, 2025. anthropic.com/...claude-sonnet-4-5
18. ^1 ^2Google. "Gemini thinking." Gemini API Documentation. ai.google.dev/...thinking
19. ^1 ^2DeepSeek. "Reasoning Model (deepseek-reasoner)." DeepSeek API Documentation. api-docs.deepseek.com/...reasoning_model
20. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9Chen, Yanda; Benton, Joe; et al. "Reasoning Models Don't Always Say What They Think." Anthropic Alignment Science, April 2025. anthropic.com/...reasoning-models-dont-say-think
21. ^arXiv. "Reasoning Models Don't Always Say What They Think." arXiv:2505.05410, May 8, 2025. arxiv.org/...2505.05410
22. ^1 ^2 ^3MarkTechPost. "Anthropic's Evaluation of Chain-of-Thought Faithfulness." April 5, 2025. marktechpost.com/...ansparency-in-reasoning-models
23. ^Anthropic. "On the Biology of a Large Language Model." Transformer Circuits, 2025. transformer-circuits.pub/...biology
24. ^1 ^2Simon Willison. "Claude 3.7 Sonnet, extended thinking and long output." Simon Willison's Weblog, February 25, 2025. simonwillison.net/...llm-anthropic-014
25. Anthropic. "Introducing Claude Haiku 4.5." Anthropic Newsroom, October 15, 2025. anthropic.com/...claude-haiku-4-5
26. Anthropic. "Introducing Claude Opus 4.5." Anthropic Newsroom, November 24, 2025. anthropic.com/...claude-opus-4-5
27. ^Anthropic. "All models overview" (archived April 2025). Anthropic API Documentation. web.archive.org/...all-models
28. ^1 ^2 ^3 ^4Anthropic. "Thinking." Claude API Documentation. Accessed November 2026. platform.claude.com/...thinking