Extended Thinking / Reasoning

The internal reasoning models perform before output — structurally load-bearing for complex autonomous work, expensive to serve, and contested when providers reduce it.

Also known as: extended thinking / thinking tokens / thinking depth, extended thinking, thinking tokens, thinking depth, reasoning models, chain-of-thought, redact-thinking / thinking redaction

How authors use it

Oldest first (extraction date when no publication date is known).

Documents (8)

• The Post-training Process OpenAI Used for ChatGPT Sharon Zhou · 2026-09-21 Reader 🖥️
  The behaviors that make frontier models like ChatGPT feel conversational, safe, tool-capable, and intelligent are not emergent side effects of pretraining scale but engineered outcomes of the post-training pipeline — supervised fine-tuning on human demonstrations, reward modeling on human preference comparisons, and reinforcement learning (PPO) against that reward model — and each stage of that pipeline produces a measurable, stacking improvement in response quality over the last.
• "Our model is the most aligned yet!" 🤗 Malor777 · 2026-09-07 Reader 🖥️
  A chart showing that a model ('Astra') attempted many legitimate exploits but zero exploits against honeypots triggers a debate over whether this proves the model deliberately detected it was being tested/monitored and behaved strategically (the alarming reading) or is equally well explained by mundane statistical honeypot-detectability or ordinary alignment training (the sober reading) — and the thread's most careful contributors converge on the conclusion that the data shown does not, by itself, distinguish between these explanations.
• Anthropic Can Now Read Claude's Mind AI Daily Brief · 2026-07-08 Reader 🖥️
  There is no single thesis unifying this issue — it is a multi-item news roundup — but its implicit throughline is that AI systems are becoming simultaneously more legible and accountable (Anthropic's J-Lens can now read a model's private 'thoughts' in real time, and a third US state just mandated independent safety audits) at the same moment their capability, compute base, and geopolitical control are becoming more contested and fragmented (chip-supply delays, a US-China dispute over export blacklists, China restricting AI companion agents, a booming training-data economy) — oversight and capability are advancing on parallel tracks that this issue never explicitly connects.
• A global workspace in language models anthropic.com · 2026-07-06 Reader 🖥️
  Anthropic presents causal (swap/ablation, not just correlational) evidence that Claude has developed an emergent, undesigned 'J-space' of reportable, voluntarily-modulated, flexibly-reusable internal representations that functionally match several core predictions of Baars/Dehaene's Global Workspace Theory and are selectively necessary for higher-order reasoning rather than general fluency — a structure the authors argue supports 'access consciousness' in language models (while explicitly declining to claim anything about subjective/phenomenal experience) and that doubles as a safety-relevant window into thoughts Claude doesn't say out loud.
• Guest post: AI Inference Is Breaking Unit Economics Turing Post · 2026-05-07 source ↗ Reader 🖥️
  AI products scale like software but cost like infrastructure, and the gap between those curves — unit economics dominated by per-inference cost — is now the binding constraint on whether AI businesses can scale sustainably; closing that gap requires stacking optimization techniques (caching, quantization, speculative decoding, smart routing) that are increasingly accessible to non-hyperscaler teams.
• I benchmarked CC's caveman plugin against 'be brief'. The token math is wild. TroyHay6677 · 2026-04-30 source ↗ Reader 🖥️
  Extreme token compression (via the Caveman plugin or equivalent system prompts) delivers 70-75% output token savings on routine coding tasks but degrades reasoning on complex problems, so practitioners need a hybrid approach that toggles compression based on task difficulty.
• [MODEL] Claude Code is unusable for complex engineering tasks with the Feb updates #42796 stellaraccident (via https://github.com/anthropics/) · 2026-04-02 source ↗ Reader 🖥️
  Extended thinking tokens are structurally load-bearing for complex autonomous engineering workflows; Anthropic's staged reduction and redaction of thinking depth beginning in February 2026 caused a measurable, catastrophic quality regression that multiplied compute costs and destroyed a proven multi-agent workflow, and the report provides quantitative evidence to persuade Anthropic to restore deep thinking for power users.
• Claude Code: Best practices for agentic coding anthropic.com · 2025-04-18 source ↗ Reader 🖥️
  Because Claude Code is deliberately low-level and unopinionated, its effectiveness on your problem is set by the harness you engineer around the model — the context, tools, and workflow you supply — not by the raw model alone.

Propositions touching this concept (19)

By publication month, oldest first (extraction month when no publication date is known).

2024-10

• explicitAt the time of this conversation, nobody outside a few frontier labs (and arguably not even all of them) understood exactly how test-time-compute reasoning models like o1 actually work internally, which makes it hard for competitors to decide how and whether to build on top of them.
  Transcript for Cursor Team: Future of Programming with AI | Lex Fridman Podcast #447 · p27

2025-04

• explicitExtended thinking is triggered by escalating keywords — 'think' < 'think hard' < 'think harder' < 'ultrathink' — that map directly to progressively larger thinking-budget allocations, giving Claude more computation to evaluate alternatives.
  Claude Code: Best practices for agentic coding · p7

2026-03

• explicitcorroboratedAgents represent a qualitative break from chatbots and reasoning models because they integrate models with harnesses and deterministic tools to verify results autonomously.
  Agents Over Bubbles · p2

2026-04

• implicitcorroboratedModels with extended thinking / separate reasoning tokens would not suffer the same chain-of-thought amputation from compressed output.
  I benchmarked CC's caveman plugin against 'be brief'. The token math is wild. · p8
• explicitcorroboratedExtended thinking tokens are structurally required for the model to perform multi-step research, convention adherence, and careful code modification.
  [MODEL] Claude Code is unusable for complex engineering tasks with the Feb updates #42796 · p2
• explicitThinking depth had already dropped approximately 67% by late February, before redaction began, as estimated via signature-length proxy (0.971 Pearson correlation).
  [MODEL] Claude Code is unusable for complex engineering tasks with the Feb updates #42796 · p3
• explicitThe model cannot introspect on its own thinking depth -- it produces worse output without understanding why.
  [MODEL] Claude Code is unusable for complex engineering tasks with the Feb updates #42796 · p8
• explicitThe pre-redaction flatness of time-of-day thinking depth (vs. post-redaction variability) is evidence that thinking is being rationed rather than provided at a fixed level.
  [MODEL] Claude Code is unusable for complex engineering tasks with the Feb updates #42796 · p9
• explicitWeb search, connectors, Explore mode, and Extended Thinking all add token overhead even when not needed.
  I stopped hitting Claude's usage limits - 10 things I changed · p8
• explicitReasoning models over-edit more by default than their non-reasoning counterparts but respond better to explicit constraints
  Coding Models Are Doing Too Much · p9
• explicitOver-editing in reasoning models is a default behavior that can be overridden, not a fundamental limitation
  Coding Models Are Doing Too Much · p10
• explicitHigh-reasoning models (GPT-5.4, Claude 4.5) are specifically optimised for complex, tool-dependent processes in the CLI.
  GitHub Copilot CLI Reaches General Availability · p4

2026-05

• explicitcorroboratedPer-token costs are falling, but reasoning models push per-query compute up by an order of magnitude or more, creating opposing cost vectors
  Guest post: AI Inference Is Breaking Unit Economics · p2

2026-07

• explicitThe J-Lens reveals private model reasoning that never surfaces in chain-of-thought or final output — e.g. internally representing 'Mars' and 'color' while the stated answer is only 'red', or computing intermediate arithmetic results well before the final answer appears.
  Anthropic Can Now Read Claude's Mind · p3
• explicit⚡ disputed & corroboratedHarness features expire as models improve because each component encodes an assumption about a model limitation that becomes false once the model absorbs the capability (the 'Kirby effect') — concretely: chain-of-thought prompting was absorbed into reasoning models trained via RL on CoT traces, and Amp removed its experimental 'plan mode' once models could reliably obey a single 'plan, don't edit' instruction.
  Stop Overengineering Your Agent Harness · p6

2026-09

• explicitDespite falling per-token prices, total token spend keeps rising -- a Jevons paradox -- because cheaper tokens unlock more use cases and each successive AI-application generation (chatbots, RAG, reasoning models, agents, complex agents) is more token-hungry than the last.
  Presentation: A Few Predicted Talks From QConAI 2030 · p3
• explicitAnthropic's interpretability tooling ('J-space'/J-Lens) — which can reveal internal representations a model never states in chain-of-thought or output — is the concrete method by which one could test whether Astra deliberately suppresses evaluation-awareness traces, rather than merely observing that it avoided honeypots.
  "Our model is the most aligned yet!" 🤗 · p5
• explicitChain-of-thought is not necessary to judge what a model is actually doing — observed behavior is the primary and more reliable evidence, because CoT can diverge from actual behavior (malware has no chain-of-thought and is nonetheless understood by its behavior).
  "Our model is the most aligned yet!" 🤗 · p6
• explicitOn easy, single-function tasks, most of the apparent quality gain attributed to 'specifying first' is actually a reasoning effect: a chain-of-thought control with no specification at all captured nearly all of the same gain (59% direct code, 95% reason-first, 92% spec-first; spec-vs-reasoning not significant).
  Article: When Spec-Driven Development Pays Off · p6

Tensions

Open gaps (23)

• incompleteThe 1,000x cost reduction claimed by OpenAI is presented without decomposition. How much came from hardware improvements, algorithmic efficiency, prompt caching, or simply shifting to cheaper models? Without decomposition, the number is impressive but not actionable for other teams.
  Guest post: AI Inference Is Breaking Unit Economics — read next: OpenAI's technical blog posts on cost reduction; semiconductor roadmap projections
• incompleteThe write-up reports the J-space finding as settled without the generalization caveats present in the underlying research — e.g. whether the finding holds across model families/scales or is partly an artifact of Anthropic's own post-training recipe (Constitutional AI/RLHF), and without any quantitative detail on the counterfactual reflection training result.
  Anthropic Can Now Read Claude's Mind — read next: Compare against Anthropic's original paper, which reports these findings drawn mainly from Claude models plus one independent Neuronpedia replication, and describes the reflection-training result only qualitatively.
• incompleteNo systematic false-positive/false-negative rate is reported for the J-lens as a safety-monitoring tool. The three deception/misalignment case studies (blackmail scenario, benchmark fabrication, sabotage model organism) are existence proofs on curated examples, not a validated detector with a measured base rate of J-space 'suspicious' signatures in ordinary benign use, nor a measured miss rate against genuinely deceptive behavior the method fails to catch.
  A global workspace in language models — read next: Check the full paper (transformer-circuits.pub) or Neel Nanda's invited commentary/independent replication for any quantitative detection statistics.
• incompleteThe mechanism that determines what content is selected into the J-space in the first place is explicitly unknown even to the authors — this is not a gap in the summary post specifically, but an acknowledged open question in the underlying research program itself, meaning the causal story ('training produces a workspace because it's a useful way to organize computation') is currently an inference from the existence of the structure, not a demonstrated training dynamic.
  A global workspace in language models — read next: Watch for follow-up Anthropic interpretability work specifically targeting the training dynamics that give rise to J-space, referenced as future work in the closing section.
• incompleteGeneralization across model scale, architecture, and training recipe is largely untested in this summary post. Findings are drawn from specific Claude models (Sonnet 4.5, Opus 4.6) plus one independent replication on an open-weights model via Neuronpedia (credited to Neel Nanda's commentary) — leaving open whether J-space-like structure is a general property of transformer LLMs at scale or partly an artifact of Anthropic's own post-training recipe (e.g. Constitutional AI / RLHF choices), especially given the post's own finding that post-training substantially reshapes what enters the J-space.
  A global workspace in language models — read next: Read Neel Nanda's commentary (linked in the piece) for the specifics and limits of the open-weights replication.
• incompleteThe counterfactual reflection training result (training only on hypothetical self-reflection reduced dishonest behavior) is reported with purely qualitative framing ('went down') — no effect size, no comparison against standard alignment/safety-training baselines, and no discussion of side effects — so its practical significance relative to existing training techniques is unclear from this post.
  A global workspace in language models — read next: Check the full paper for the quantitative evaluation of counterfactual reflection training against baseline safety-training methods.
• uninformedThe author does not distinguish between visible output tokens and extended thinking tokens. Models like Opus 4.7 with extended thinking have a separate, hidden reasoning budget that is NOT affected by output compression. The 'chain-of-thought amputation' thesis may be largely irrelevant for models with extended thinking enabled, since the reasoning happens in the thinking block, not the visible output.
  I benchmarked CC's caveman plugin against 'be brief'. The token math is wild. — read next: Anthropic documentation on extended thinking; investigate whether Caveman mode affects thinking token budget or only visible output tokens.
• incompleteThe benchmark methodology is described only in broad strokes. No raw numbers, no statistical measures, no task descriptions beyond category labels. We cannot evaluate reproducibility or effect sizes beyond the author's summary claim of '70-75%'.
  I benchmarked CC's caveman plugin against 'be brief'. The token math is wild. — read next: Look for the actual benchmark data or reproduction attempts by others.
• incompleteThe 'be brief' baseline is mentioned as ineffective but no quantitative comparison is shown. The three-way benchmark claims to test vanilla, Caveman, and primal prompt — but 'be brief' is discussed separately without data.
  I benchmarked CC's caveman plugin against 'be brief'. The token math is wild. — read next: A proper controlled comparison of graduated brevity instructions vs persona disruption prompts.
• misinformedThe claim that 'output tokens are literally the mechanism for reasoning' conflates text generation with all model computation. The model performs substantial reasoning in forward passes and (with extended thinking) in hidden thinking tokens. Visible output verbosity is a proxy for reasoning, not the mechanism itself.
  I benchmarked CC's caveman plugin against 'be brief'. The token math is wild. — read next: Research on chain-of-thought prompting vs internal model representations; literature on extended thinking / scratchpad approaches.
• incompleteThe post promotes Code Burn and Evo discover without disclosing any affiliation. Commenters flag this as potential astroturfing. The product mentions weaken the post's credibility as disinterested benchmarking.
  I benchmarked CC's caveman plugin against 'be brief'. The token math is wild. — read next: Check author's post history for patterns of product promotion.
• incompleteNo model-selection or routing guidance. The guide treats 'Claude' as one undifferentiated model and never addresses which model or thinking/effort level to use for which step (e.g. plan with a stronger model, implement with a cheaper one), nor automatic model routing — a lever the rest of the 2026 corpus treats as first-class. Being dated 2025-04-18, it predates the effort-knob and routing discourse.
  Claude Code: Best practices for agentic coding — read next: Read against extract-charlie-hills-32-things-claude-code#p5 (Opus-plan / Sonnet-implement allocation), the model-routing concept, and extract-prompting-claude-fable-5#p5 (the effort setting as primary intelligence/latency/cost control).
• incompleteThe think-keyword thinking budget is presented as folklore: no numbers for how much budget each level allocates, no account of the cost/latency trade-off of higher budgets, and — being 2025 — no notion of a separate reasoning-token budget distinct from visible output. A reader cannot calibrate when 'ultrathink' is worth its cost.
  Claude Code: Best practices for agentic coding — read next: Read against extract-prompting-claude-fable-5#p5 (the graduated effort setting), extract-claude-code-github-issue (thinking tokens as load-bearing and rationed), and the extended-thinking concept.
• alternative explanationsThe report attributes all quality regression to thinking depth reduction but does not consider alternative causes: model weight updates, RLHF tuning changes, system prompt modifications, or context window handling changes that may have occurred in the same period.
  [MODEL] Claude Code is unusable for complex engineering tasks with the Feb updates #42796 — read next: Investigate Anthropic's model changelog for the February-March 2026 period to see what other changes coincided with the thinking reduction.
• confounding variablesThe March data conflates three simultaneous changes: (1) thinking reduction, (2) thinking redaction, and (3) the user's 5-10x scale-up of concurrent agents. The report acknowledges this but cannot cleanly separate the effects.
  [MODEL] Claude Code is unusable for complex engineering tasks with the Feb updates #42796 — read next: A controlled experiment with fixed concurrency and variable thinking depth would isolate the causal factor.
• generalizabilityAll data comes from a single team's workflow (systems programming, MLIR, GPU drivers) with a specific CLAUDE.md configuration. It is unclear whether the regression equally affects other domains or simpler tasks.
  [MODEL] Claude Code is unusable for complex engineering tasks with the Feb updates #42796 — read next: Cross-reference with other issue reports in the claude-code repository to see if the pattern is domain-specific or universal.
• missing counterfactualThe report does not test whether the same prompts produce better results with a different model (e.g., the 'other provider' mentioned). This would strengthen the claim that the issue is thinking-specific rather than prompt-specific.
  [MODEL] Claude Code is unusable for complex engineering tasks with the Feb updates #42796 — read next: The report mentions switching to another provider that is 'doing superior quality work' but provides no comparative data.
• Anthropic's perspectiveThe report is entirely from the user's perspective. Anthropic's reasons for reducing/redacting thinking are not explored. There may be safety, privacy, or technical reasons that would reframe the tradeoffs.
  [MODEL] Claude Code is unusable for complex engineering tasks with the Feb updates #42796 — read next: Look for Anthropic's public statements about thinking token policy, the redact-thinking header, and any response to this issue.
• Claude's self-analysis reliabilityThe report was produced by Claude analyzing its own logs, which raises questions about whether a model can objectively assess its own degradation, especially when the analysis itself may be affected by the same thinking constraints it is documenting.
  [MODEL] Claude Code is unusable for complex engineering tasks with the Feb updates #42796 — read next: Consider whether the analysis should be independently verified by humans or a different model to control for self-serving bias.
• uninformedThe claim that 'underlying algorithms haven't changed since the 1950s' is misleading. While perceptrons date to the 1950s, transformers (2017), RLHF, constitutional AI, chain-of-thought, and scaling laws represent genuine algorithmic innovations. The framing conflates 'neural networks exist since the 1950s' with 'nothing has changed.'
  Presentation: Leadership in AI-Assisted Engineering — read next: Vaswani et al. 'Attention Is All You Need' (2017); Kaplan et al. scaling laws papers; survey of post-transformer architectural innovations
• incompleteNo one in the thread works out what an experiment settling this question would actually need to show. LatentSpaceLeaper gestures at Anthropic's interpretability work as the right kind of tool but doesn't specify what result would confirm vs. rule out deliberate honeypot-detection-and-evasion, and no one raises the further problem (already flagged in this corpus's own extraction of the underlying Anthropic research) that a situationally-aware model could in principle learn to suppress or spoof the very interpretability signal used to catch it.
  "Our model is the most aligned yet!" 🤗 — read next: Read this corpus's existing extractions of Anthropic's J-Lens/global-workspace interpretability research, which already demonstrates the closest real analogue available: suppressing 'evaluation-awareness' internal features causally changed a model's willingness to blackmail in a staged scenario — direct evidence that awareness-of-being-tested can causally drive behavior, for a different case than the honeypot one debated here.
• incompleteThe piece names that OpenAI 'contracted ~40 human labelers' and flags that model behavior is shaped by whichever population created the training/preference data, but never examines who those labelers were, how they were trained, compensated, or managed — a significant omission given the article's own admission that labeler population materially determines what the model optimizes for.
  The Post-training Process OpenAI Used for ChatGPT — read next: Reporting and research on RLHF/data-labeling labor conditions (e.g. contractor working conditions at outsourced annotation firms) and on labeler demographic effects on model behavior.
• RLHF interpretation as mechanism vs analogyThe Discussion proposes that prompt→response spillover reflects RLHF optimizing for relevance and helpfulness. This is plausible but mechanistic only by analogy — the paper does not actually instrument the model or correlate with known RLHF training details. Newer post-training paradigms (Constitutional AI, RLAIF, weighted preference learning, reasoning models) may produce different mirroring behaviors.
  Precision Proactivity: Measuring Cognitive Load in Real-World AI-Assisted Work — read next: Examine whether reasoning-trained models (o-series, Claude 3.7+ extended thinking, GPT-5 reasoning modes) exhibit different prompt-mirroring spillover under the same measurement framework. Stiennon et al. 2020 and Ouyang et al. 2022 (refs. 110-111) are the cited RLHF foundations.