---
title: "Claude Code Effort Levels: Is Medium on Opus 5.5 Already the Default?"
date: "2026-10-13"
video_id: "7696113861551738144"
tiktok_url: "https://www.tiktok.com/@mattpocockuk/video/7696113861551738144"
creator: "@mattpocockuk"
category: "AI"
tags: ["claude code", "effort level", "opus 5.5", "extended thinking", "context window"]
excerpt: "Claude Code's effort levels explained: low to max, the per-model defaults, why medium is already Opus 5.5's default, and when lowering effort saves tokens."
---
Matt Pocock (@mattpocockuk) tells viewers in a TikTok to "freaking lower your effort," and says he runs Opus 5.5 on medium while debating a move to low. His point about overthinking matches Anthropic's own documentation. But medium is already the shipped default on Opus 5.5, so running it there is the baseline rather than a reduction. Only the move to low he is considering would lower effort.

## What are Claude Code's effort levels?

The [Claude Code model configuration docs](https://code.claude.com/docs/en/model-config) list five levels: `low`, `medium`, `high`, `xhigh` and `max`. The video's caption writes the fourth as "X high," which is how it sounds.

Support varies by model. Fable 5.1, Fable 5, Opus 5.5, Sonnet 5.5, Haiku 5.5, Opus 5, Sonnet 5, Opus 4.8 and Opus 4.7 accept all five. Opus 4.6 and Sonnet 4.6 accept `low`, `medium`, `high` and `max`, without `xhigh`.

The API uses the same names under `output_config.effort`. The [API effort page](https://platform.claude.com/docs/en/build-with-claude/effort) warns: "Don't pass `adaptive` as an `effort` value: `adaptive` is a thinking mode, not an effort level."

## What is the default effort level on Opus 5.5?

Medium. The model configuration page says `high` is the "Default on every model that supports effort," with exceptions: "`medium`: Default on Opus 5.5, Sonnet 5.5, and Haiku 5.5" and "`xhigh`: Default on Opus 4.7."

The API docs agree: "Claude Opus 5.5 and Claude Haiku 5.5 default to medium," and setting the default "produces exactly the same behavior as omitting the `effort` parameter entirely."

Anthropic gives a reason for the choice: "In Anthropic's testing, Opus 5.5 at `medium` matches or exceeds Opus 5 at `high` on coding and knowledge-work evaluations." The [Opus 5.5 launch page](https://www.anthropic.com/claude-opus-5-5), dated 22 September 2026, says "Opus 5.5 at default effort beats Opus 5 at max effort for about a fifth of the cost" on Terminal-Bench 4.0.

So an Opus 5.5 user who never touches the setting is already on medium.

## Does high effort make Claude overthink?

At the top of the scale, yes, according to Anthropic. The model configuration page says "`max` may show diminishing returns and is prone to overthinking, so test before adopting it broadly." The API guidance for Opus 4.7 says that "On most workloads `max` adds significant cost for relatively small quality gains, and on some structured-output or less intelligence-sensitive tasks it can lead to overthinking."

Claude Academy's [effort tutorial](https://academy.claude.com/tutorials/choosing-the-right-effort-level-in-claude-code) describes the symptom: "When the effort is too high...you will notice a pattern of overthinking, or of identifying more tasks than you actually assigned." It recommends a narrower method than a blanket cut: "Start at the default, watch what comes back, and adjust one level at a time."

## When should you use low effort?

Pocock's examples are small code changes and repo exploration. The docs describe `low` as fitting "Quick exchanges where you review each result, such as brainstorming, a first sketch, or a small change like a rename." The API page calls it "Most efficient. Significant token savings with some capability reduction," suited to "Simpler tasks that need the best speed and lowest costs, such as subagents."

The [cost docs](https://code.claude.com/docs/en/costs) also suggest you "Delegate verbose operations to subagents," so verbose output stays in the subagent's context and only a summary returns.

One caveat: the API page says effort is "not a strict token budget." At lower levels, "Claude still thinks on sufficiently difficult problems, but thinks less." Dropping to low makes fewer tokens likely without guaranteeing them.

## How to change the effort level

In Claude Code, the level comes from `CLAUDE_CODE_EFFORT_LEVEL`, the `--effort` launch flag, or `/effort` during a session, then from a per-model `modelSettings` entry or the `effortLevel` settings key, then from the model default. Skills and subagents accept an `effort:` key in frontmatter. In the API it is `output_config.effort`. The video does not name any of these.

## Does lower effort keep more room in the context window?

Indirectly. Pocock says "The fewer tokens you use, the smarter the model remains and the more juice you have in your context window." The API page says effort "affects **all tokens** in the response," including text, tool calls and thinking, and "Lower effort also means fewer and terser tool calls." It controls output, not context directly.

The context half is real. Anthropic's [context engineering post](https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents) from 29 September 2025 says "as the number of tokens in the context window increases, the model's ability to accurately recall information from that context decreases," and calls context "a finite resource with diminishing marginal returns." Because "Claude Code sends your full conversation with every request," fewer output tokens per turn slow how fast the context grows.

On Opus 5.5, effort is also the only lever. The cost docs say "You can't turn off thinking on Opus 5.5, Sonnet 5.5, Haiku 5.5, or the Fable models," and the API returns a 400 error for `thinking: {"type": "disabled"}` on Opus 5.5.

## Bottom line

Pocock is right that high, xhigh and max are more than most coding tasks need, and Anthropic says max is prone to overthinking. On Opus 5.5, medium is the default, so lowering effort means choosing low for simple work like renames, small edits and exploration subagents. Adjust one level at a time and watch the results.

## Sources

- [Model configuration, Claude Code Docs](https://code.claude.com/docs/en/model-config) - effort levels, per-model defaults and settings
- [Effort, Claude Platform Docs](https://platform.claude.com/docs/en/build-with-claude/effort) - API parameter, defaults and scope
- [Manage costs effectively, Claude Code Docs](https://code.claude.com/docs/en/costs) - thinking, subagents and context growth
- [Choosing the right effort level in Claude Code, Claude Academy](https://academy.claude.com/tutorials/choosing-the-right-effort-level-in-claude-code) - overthinking symptoms and tuning method
- [Effective context engineering for AI agents, Anthropic](https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents) - context rot and attention budget
- [Introducing Claude Opus 5.5, Anthropic](https://www.anthropic.com/claude-opus-5-5) - launch date, pricing and default effort
- [mattpocock/skills on GitHub](https://github.com/mattpocock/skills) - Pocock's published skills
- [AI Hero](https://www.aihero.dev/) - the sponsor shown in the video


## Video description and verification notes

**A 74.30-second vertical clip, 1080x1920 at 60 fps, delivered as a QuickTime/MP4 container (`mov,mp4,m4a,3gp,3g2,mj2`) holding an HEVC Main profile video stream tagged `hvc1` in `yuv420p` at 670,136 bps across 4,455 frames, and an AAC HE-AACv2 audio stream at 44,100 Hz stereo, 64,123 bps across 1,603 frames; overall bitrate 742,710 bps, filesize 6,897,922 bytes, format id `bytevc1_1080p_742710-1`, dynamic range SDR. `ffprobe` reports the video stream at 74.250000 s and the audio stream and container at 74.300000 s. Posted at epoch `1791891131`, which is 2026-10-13T11:32:11Z, by `mattpocockuk` (nickname "Matt Pocock", uploader id `6822980578351973381`), with audio credited to "original sound - Matt Pocock" and `artist` set to "Matt Pocock". The description is 35 characters and is a verbatim copy of the title, "Your effort level is STILL TOO HIGH". At the capture time recorded on `metadata.json` (epoch `1792387293`, file mtime 2026-10-19 00:21:33 local), yt-dlp recorded 27,000 views, 685 likes, 27 comments, 32 reposts, and 179 saves. I read 15 of the 37 extracted frames: 001, 002, 003, 004, 005, 006, 010, 016, 022, 025, 026, 027, 035, 036, and 037, spanning 0 s to 72 s at 2-second intervals. The transcript runs 276 words across 15 segments. Every frame I read shows the same shot: a bald man in a white t-shirt and dark round tortoiseshell glasses, seated in a black office chair with a visible headrest logo, speaking into a large black cardioid microphone on a desk boom with a blue-collared XLR cable, against a dark green wall flanked by two pine bookcases holding blue and teal hardbacks. Frame 001 (0 s) carries two overlays: a burned-in yellow monospace caption reading "You're going to make me say" and a lower-third sponsor card with an orange, blue and yellow gradient border, black inner panel, and the text "AIhero.dev" above "Go from Zero To AI Hero". That card is already gone by frame 002 (2 s), which shows only the caption "it again, aren't you? The effort level". Captions continue through the clip: frame 003 (4 s) "is way too high. People just go", frame 004 (6 s) "nuts. They think, I need to make", frame 005 (8 s) "sure that my model is putting in", frame 006 (10 s) "is the most important to me. But", frame 010 (18 s) "Minecraft sculpture. / The more effort you put", frame 016 (30 s) "what is more effort going to", frame 022 (42 s) "doesn't have to / burn. Because with", frame 025 (48 s) "so when you've got X high,", frame 026 (50 s) "when you've got max, even if", frame 027 (52 s) "you've got high, I still think", frame 035 (68 s) "actually doing / useful tasks. And so", frame 036 (70 s) "Opus 5.5 on medium and I'm", and frame 037 (72 s) "debating even moving to low." No terminal, editor, settings file, `/effort` slider, token counter, benchmark chart, or documentation page appears in any frame I read; the clip is a continuous talking-head take with no screen recording and no demonstration of the setting it argues about. The spoken argument is that most people run their coding agent at too high an effort level, that effort behaves like detail in a drawing or a Minecraft build and therefore buys nothing on simple tasks such as small edits and repo exploration, that high effort makes the model overthink and burn context, and that the speaker personally runs Opus 5.5 at medium while considering low. The last transcript segment ends at 74.14 s against a measured audio duration of 74.30 s, so nothing in this transcript is timestamped past the end of the audio and I treat all 15 segments as real.**

## The named effort levels are real, and the docs spell them exactly as the captions do

At 49.06 s the video says "when you've got X high, when you've got max, even if you've got high, I still think that's way too much effort for most coding tasks." The burned-in caption in frame 025 renders the first of those as "X high", two words, which is a transcription of how the term sounds rather than how it is written.

The Claude Code model configuration documentation lists five levels and writes them as single lowercase tokens: `low`, `medium`, `high`, `xhigh`, and `max`. The per-model support table states that "Fable 5.1 and Fable 5" and "Opus 5.5, Sonnet 5.5, Haiku 5.5, Opus 5, Sonnet 5, Opus 4.8, and Opus 4.7" all accept `low`, `medium`, `high`, `xhigh`, `max`, while "Opus 4.6 and Sonnet 4.6" accept `low`, `medium`, `high`, `max` with no `xhigh`.

The API uses the same five names. The Claude Platform effort page instructs you to "Set `output_config.effort` on the request" and gives the level values as `max`, `xhigh`, `high`, `medium`, and `low`. The page also warns against one value the video does not mention: "Don't pass `adaptive` as an `effort` value: `adaptive` is a thinking mode, not an effort level."

So the three levels the video names at 49.06 s all exist, and the ordering it implies, with `max` above `xhigh` above `high`, matches the documented ordering. The written form is `xhigh`.

## "I use Opus 5.5 on medium" is the shipped default, not a step down

At 68.52 s the video says "And so freaking lower your effort. I use Opus 5.5 on medium, and I'm debating even moving to low." Frame 036 at 70 s burns the same claim into the caption track as "Opus 5.5 on medium and I'm".

The Claude Code model configuration documentation gives the per-model defaults directly. `high` is "Default on every model that supports effort", with named exceptions: "`medium`: Default on Opus 5.5, Sonnet 5.5, and Haiku 5.5" and "`xhigh`: Default on Opus 4.7". The API side matches. The Claude Platform effort page states that "Most Claude models default to high effort, spending as many tokens as needed for excellent results; Claude Opus 5.5 and Claude Haiku 5.5 default to medium", and adds that "Setting `effort` to the model's default (`"medium"` on Claude Opus 5.5 and Claude Haiku 5.5, `"high"` on other models) produces exactly the same behavior as omitting the `effort` parameter entirely."

The video frames medium on Opus 5.5 as the speaker's own disciplined choice in a segment whose explicit instruction is "lower your effort." On Opus 5.5 that configuration is what you get by changing nothing. Anthropic sets it there deliberately: the model configuration page gives the reason as "In Anthropic's testing, Opus 5.5 at `medium` matches or exceeds Opus 5 at `high` on coding and knowledge-work evaluations." The Opus 5.5 launch page makes the same point against a stronger baseline, stating that "Opus 5.5 at default effort beats Opus 5 at max effort for about a fifth of the cost" on Terminal-Bench 4.0.

The only part of the sentence that would actually lower effort below the shipped baseline is the part he has not done yet, "debating even moving to low." That is the claim in the video, and frame 037 confirms the caption "debating even moving to low." at 72 s.

## The overthinking claim at max is in Anthropic's own documentation, almost in the video's words

At 54.28 s the video says high effort "is actually going to distract the model. It's going to push it towards outputs that are not actually what you want." Earlier, at 34.40 s, it says more effort on a simple task means "it's probably going to overthink things."

Anthropic documents the same failure mode at the top of the scale. The Claude Code model configuration page describes `max` as suited to "Hard problems you want Claude to work through without you, such as finding security vulnerabilities" and then states that "`max` may show diminishing returns and is prone to overthinking, so test before adopting it broadly." The per-model API guidance for Opus 4.7 is blunter: "Reserve for frontier problems. On most workloads `max` adds significant cost for relatively small quality gains, and on some structured-output or less intelligence-sensitive tasks it can lead to overthinking."

Claude Academy's tutorial on choosing an effort level describes the symptom the user sees: "When the effort is too high...you will notice a pattern of overthinking, or of identifying more tasks than you actually assigned." It also states the design intent behind the shipped defaults: "The default sits in the middle, and it is not arbitrary: it is tuned for each model to the level where most tasks finish without overspending."

That tutorial's method is more conservative than the video's instruction. It advises "Start at the default, watch what comes back, and adjust one level at a time", and says you "adjust down for tasks that you know are simple and should not require a big spend of resources: time, turns, tool calls, or document reads." The video's framing, a blanket "freaking lower your effort" at 68.52 s, skips the per-task judgment the documentation asks for.

## The exploration example matches the documented guidance, with one caveat the video leaves out

At 28.92 s the video says "if your agent is just making some small code changes, or it's just exploring the repo, what is more effort going to do?" At 44.76 s it adds "With an exploration, it's going to burn a ton of tokens that it doesn't have to burn. Because with exploration, all you've got to do is just open some files and then you're good to go."

Both examples appear in the documentation. The Claude Code model configuration page describes `low` as fitting "Quick exchanges where you review each result, such as brainstorming, a first sketch, or a small change like a rename", which is the rename-scale edit the video has in mind. The API effort page describes `low` as "Most efficient. Significant token savings with some capability reduction", with the typical use case "Simpler tasks that need the best speed and lowest costs, such as subagents." Claude Code's cost documentation routes exploration work the same way, advising you to "Delegate verbose operations to subagents" so that "the verbose output stays in the subagent's context while only a summary returns to your main conversation", and to "choose a smaller model for a subagent."

The caveat the video drops is that effort is advisory rather than a cap. The API effort page states: "Effort is a behavioral signal, not a strict token budget. At lower effort levels, Claude still thinks on sufficiently difficult problems, but thinks less than it would at higher effort levels for the same problem." Dropping to `low` for a repo exploration does not guarantee the token saving the video promises; it biases the model toward it.

The video also never names the control. In Claude Code the level resolves from `CLAUDE_CODE_EFFORT_LEVEL`, the `--effort` launch flag, or `/effort` in session, then from a per-model entry in `modelSettings` or the `effortLevel` settings key, then from the model's default. Skills and subagents take an `effort:` key in frontmatter. In the API it is `output_config.effort`. None of those strings is spoken in the transcript or visible in any frame I read.

## Effort governs more than thinking, which weakens the "smarter model" claim

At 63.76 s the video says "The fewer tokens you use, the smarter the model remains and the more juice you have in your context window for actually doing useful tasks." That sentence runs two different mechanisms together.

The API effort page is explicit about scope: "The effort parameter affects **all tokens** in the response, including: Text responses and explanations; Tool calls and function arguments; Thinking (when active)." It continues, "Because effort applies to every output token, it works whether or not thinking is enabled. Lower effort also means fewer and terser tool calls." Effort is therefore an output-token dial, not a context-window dial.

The "smarter the model remains" half has separate support. Anthropic's engineering post on context engineering, published 29 September 2025, defines the degradation the video is gesturing at: "as the number of tokens in the context window increases, the model's ability to accurately recall information from that context decreases." It frames the mechanism as a budget: "LLMs have an 'attention budget' that they draw on when parsing large volumes of context. Every new token introduced depletes this budget by some amount. Context, therefore, must be treated as a finite resource with diminishing marginal returns."

The connection between the two is real but indirect. Lower effort produces fewer output tokens per turn, and in a multi-turn agent session those outputs accumulate in the conversation that gets resent on every subsequent request, so a lower setting does slow the growth of context. Claude Code's cost page describes that accumulation independently: "Claude Code sends your full conversation with every request, and each time Claude uses tools it sends another request carrying that batch of tool results." The video asserts the link as if effort directly governed context occupancy, and it does not.

One further qualification the video omits: on Opus 5.5 you cannot escape thinking by any route other than effort. The Claude Code cost page states, "You can't turn off thinking on Opus 5.5, Sonnet 5.5, Haiku 5.5, or the Fable models, which always use extended thinking." The API effort page says the same thing in error terms for Opus 5.5: "Requests that set `thinking: {"type": "disabled"}` return a 400 error at every effort level." Effort is the only lever on that model, which makes the video's advice more load-bearing than it explains.

## What the artifact does and does not show

The clip is a single unbroken talking-head shot. Across frames 001 through 037 the framing, lighting, background, chair, microphone, and wardrobe do not change, and the only overlays are the yellow monospace caption track and the AIhero.dev sponsor card that appears in frame 001 and is gone by frame 002. There is no screen recording, no `/effort` slider, no `settings.json`, no `/usage` output, no before-and-after token count, and no benchmark figure anywhere in the 15 frames I read.

The AIhero.dev card in frame 001 reads "AIhero.dev" over "Go from Zero To AI Hero". Fetching aihero.dev returns the tagline "AI Hero teaches engineers to build with AI through courses, cohorts, events, and free tutorials." The site's own homepage copy does not name Matt Pocock, so the attribution of AI Hero to him comes from secondary sources rather than the site itself, and I am not citing it as primary.

One transcript error is worth flagging because it changes the opening line. Whisper's output begins "going to make me say it again, aren't you?", dropping the first word. The burned-in caption in frame 001 reads "You're going to make me say", so the real opening is "You're going to make me say it again, aren't you?" The video's own caption track is the better record of the first second of audio than the generated transcript is.

The sister repository the title's "STILL" implies, `mattpocock/skills`, describes itself as "Skills for Real Engineers. Straight from my .agents directory." Its listed skills include `ask-matt`, `grill-with-docs`, `triage`, `improve-codebase-architecture`, `to-spec`, `to-tickets`, `implement`, `wayfinder`, `retro`, `prototype`, `diagnosing-bugs`, `research`, `tdd`, `domain-modeling`, `codebase-design`, `code-review`, `pr`, `wizard`, `grill-me`, `handoff`, `teach`, `to-questionnaire`, `wait-what`, `grilling`, and `writing-for-agents`. None of them documents an effort level, a thinking budget, or a token budget, so the advice in this video is not reflected in the published skill frontmatter that Claude Code reads via the `effort:` key.

## Key Takeaways

- **Verified:** The effort levels the video names at 49.06 s all exist. Claude Code and the Claude API both document five levels, written `low`, `medium`, `high`, `xhigh`, and `max`. The caption's "X high" in frame 025 is the spoken form of `xhigh`.
- **Correction:** The video presents "I use Opus 5.5 on medium" at 68.52 s inside an instruction to lower your effort. Anthropic's documentation states that `medium` is the default on Opus 5.5 in Claude Code and that on the API, setting `medium` on Opus 5.5 "produces exactly the same behavior as omitting the `effort` parameter entirely." His stated configuration is the shipped baseline.
- **Verified:** The overthinking claim at 34.40 s and 54.28 s matches Anthropic's own wording. The model configuration page says `max` "may show diminishing returns and is prone to overthinking, so test before adopting it broadly."
- **Partial correction:** The claim at 63.76 s that "The fewer tokens you use, the smarter the model remains and the more juice you have in your context window" merges two mechanisms. Effort controls output tokens, including tool calls and explanations, not context occupancy. The recall degradation it describes is real and documented as context rot, and lower effort slows context growth indirectly rather than governing it.
- **Clarification:** Effort is advisory. The API effort page states it is "a behavioral signal, not a strict token budget", so dropping to `low` for a repo exploration biases toward fewer tokens rather than guaranteeing them.
- **Clarification:** On Opus 5.5 thinking cannot be disabled at any effort level, with the API returning a 400 error for `thinking: {"type": "disabled"}`. Effort is the only control, which the video does not mention.
- **Context:** Claude Academy's effort tutorial recommends a narrower method than the video's blanket instruction: "Start at the default, watch what comes back, and adjust one level at a time."
- **Context:** The video never names the setting. Claude Code exposes it as `/effort`, `--effort`, `CLAUDE_CODE_EFFORT_LEVEL`, `effortLevel`, `modelSettings`, and skill or subagent `effort:` frontmatter; the API exposes it as `output_config.effort`.
- **Context:** No frame I read shows a terminal, settings file, effort slider, token count, or benchmark. The clip is 74.30 seconds of continuous talking head with a caption track and a one-frame AIhero.dev sponsor card.
- **Correction:** The generated transcript drops the first word. Frame 001's burned-in caption reads "You're going to make me say", so the clip opens "You're going to make me say it again, aren't you?"
- **Unverified:** The video's implicit claim that most viewers are running above the default is not something I can source. No usage distribution for effort level settings is published by Anthropic.

## Resources

- [Model configuration, Claude Code Docs, code.claude.com](https://code.claude.com/docs/en/model-config) establishes the five effort level names, the per-model support table, the per-model defaults including `medium` on Opus 5.5, the resolution order across `CLAUDE_CODE_EFFORT_LEVEL`, `--effort`, `/effort`, `modelSettings` and `effortLevel`, the `max` overthinking warning, and the statement that Opus 5.5 at `medium` matches or exceeds Opus 5 at `high` in Anthropic's testing.
- [Effort, Claude Platform Docs, platform.claude.com](https://platform.claude.com/docs/en/build-with-claude/effort) establishes the `output_config.effort` parameter path, the five valid values, the default of `high` on most models and `medium` on Opus 5.5 and Haiku 5.5, the statement that setting the default produces behavior identical to omitting the parameter, the scope of effort over all output tokens, the "behavioral signal, not a strict token budget" qualification, and the 400 error for disabled thinking on Opus 5.5.
- [Manage costs effectively, Claude Code Docs, code.claude.com](https://code.claude.com/docs/en/costs) establishes that thinking cannot be turned off on Opus 5.5, Sonnet 5.5, Haiku 5.5 or the Fable models, the guidance to lower effort via `/effort` for simpler tasks, the subagent delegation advice for verbose operations, and the description of full-conversation resend that drives context growth.
- [Choosing the right effort level in Claude Code, Claude Academy](https://academy.claude.com/tutorials/choosing-the-right-effort-level-in-claude-code) establishes the overthinking symptom description, the "one level at a time" method, and the statement that each model's default is tuned to finish most tasks without overspending.
- [Effective context engineering for AI agents, Anthropic, 29 September 2025](https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents) establishes the context rot definition, the attention budget framing, and the "finite resource with diminishing marginal returns" characterization I cite against the video's 63.76 s claim.
- [Introducing Claude Opus 5.5, Anthropic, 22 September 2026](https://www.anthropic.com/claude-opus-5-5) establishes the model's announcement date, its $4 per million input and $20 per million output pricing, its `medium` default effort, and the claim that Opus 5.5 at default effort beats Opus 5 at max effort for about a fifth of the cost on Terminal-Bench 4.0.
- [mattpocock/skills, GitHub](https://github.com/mattpocock/skills) establishes the repository's self-description and its skill list, none of which sets an effort level, thinking budget, or token budget.
- [AI Hero, aihero.dev](https://www.aihero.dev/) establishes the tagline on the sponsor card shown in frame 001 and confirms the site does not name Matt Pocock in its homepage copy.
