---
title: "Claude Effort Levels: Does Max Effort Really Use 3x the Tokens?"
date: "2026-10-05"
video_id: "7693317987918499085"
tiktok_url: "https://www.tiktok.com/@forgedlabsai/video/7693317987918499085"
creator: "@forgedlabsai"
category: "AI"
tags: ["claude effort", "claude fable 5.1", "claude opus 5.5", "terminal-bench", "ai costs"]
excerpt: "Anthropic's effort test shows Claude Fable 5.1 using about 3x the tokens at max effort. Here is what improved, which domains gained most, and model costs."
---
It does in one specific test. Anthropic's own benchmark sweep shows Claude Fable 5.1 using a median of 73k tokens per attempt at low effort and 222k at max effort on Terminal-Bench 3.0, about 3.04 times as many. A TikTok summarizing that result gets the number right but picks the wrong headline domain, and it applies coding results to business tasks the benchmark never tested.

The video, from @forgedlabsai (Forged Labs), opens with "Most companies are about to waste big money by asking AI to think deeply about everything." Its advice is to match AI effort to the cost of being wrong.

## Where does the 3x token figure come from?

The chart in the video comes from [Spending your effort](https://claude.dev/blog/spending-your-effort/), published on claude.dev on September 25, 2026 by Thariq Shihipar. The site describes itself as "technical writing for people building with Claude," sharing "POVs from Anthropic's developers." Shihipar's bio on ai.engineer lists him as a "member of technical staff at Anthropic working on Claude Code," so the video's "anthropic engineer" description is correct.

The article reports Claude Fable 5.1 at low effort as "370 attempts, median 73k tokens each" and at max effort as "370 attempts, median 222k tokens each." That ratio is 3.04. The 370 attempts match the benchmark: Terminal-Bench 3.0 shipped in July 2026 with 74 tasks, and 74 tasks at five attempts each is 370.

The video leaves out the scope. The figure covers one model, Claude Fable 5.1, at the two ends of a five-level scale (low and max), on one benchmark. The chart plots four models with four separate curves, and the narration names none of them.

## Which tasks benefit most from higher effort?

The video says higher effort helped "especially security." The article's domain table shows pass rates at low and max effort:

| Domain | Low | Max | Change |
|---|---|---|---|
| Security | 64% | 87% | +23 points |
| Hardware | 34% | 75% | +41 points |
| ML | 54% | 73% | +19 points |
| Science | 41% | 61% | +20 points |
| Software | 43% | 56% | +13 points |
| Media | 18% | 30% | +12 points |
| Operations | 12% | 22% | +10 points |

Security improved, but hardware gained the most, at 41 points against security's 23. Hardware more than doubled its pass rate.

The security framing does hold for one example in the article. A task called `html-js-filter` asks for "an HTML sanitizer that strips every way of smuggling JavaScript into a page," and "Fable 5.1 went from 1/5 at low to 5/5 at xhigh." That is one task, not a domain ranking.

## Does higher effort help less on routine tasks?

The video says higher effort "delivered much smaller gains on some routine tasks." In percentage points, operations (+10) and media (+12) had the smallest gains, so the claim holds on that measure.

Two details complicate it. The article never calls those domains routine. They are the two where the models score worst overall, and operations reaches only 22% even at max effort. In relative terms, operations rose from 12% to 22%, an 83% increase, and media from 18% to 30%, a 67% increase. Both are larger relative gains than security's move from 64% to 87%.

## How Anthropic suggests choosing an effort level

The article's guidance is written for coding:

- **Low:** "for when I want quick responses that are in the loop, e.g. brainstorming, sketching, easy changes."
- **Medium:** "for most of my regular software engineering work, e.g. new feature implementation."
- **High:** "for work where verification is important or there are edge cases, e.g. fixing a bug in a brownfield codebase."
- **Max:** "When I want Claude to operate fully autonomously to solve difficult problems, e.g. end to end building and verification of an app."

## Does this apply to finance and compliance work?

The video extends the advice to "financial analysis," "customer commitments," "compliance and final verification." None of those are in the source.

[Terminal-Bench](https://www.tbench.ai/) describes itself as "a benchmark to measure and evolve with the frontier of agent work," hosted by Stanford, Harbor and the Laude Institute. Snorkel AI describes it as testing whether agents can operate real terminals, "writing and debugging code, configuring environments, and recovering from failure, all without a GUI." The article's domains are hardware, science, ML, software, security, operations and media. There is no finance, compliance or customer-commitment task.

The extension to business work may be a reasonable analogy, but it is not a finding from this data. The opening claim about companies wasting "big money" comes with no spend figure or source.

## Is a cheaper model a better lever than effort?

The chart carries an annotation the narration never mentions: "same score, half the tokens." It sits beside the Opus 5.5 "high" point, level with the Fable 5.1 "max" point at about twice the token count. The article says "Fable 5.1 and Opus 5.5's effort curves are our best yet: at each level, there is an uptick in benchmark scores and tokens consumed."

Pricing widens the gap. Per [Anthropic's pricing page](https://platform.claude.com/docs/en/about-claude/pricing):

| Model | Input per MTok | Output per MTok |
|---|---|---|
| Claude Opus 5.5 | $4 | $20 |
| Claude Opus 5 | $5 | $25 |
| Claude Fable 5.1 | $10 | $50 |
| Claude Fable 5 | $10 | $50 |

Opus 5.5 costs 40% of Fable 5.1 per token. Anthropic's Opus 5.5 launch page says "Opus 5.5 at default effort beats Opus 5 at max effort for about a fifth of the cost." Following only the video's advice means adjusting effort while leaving model choice alone, and on this chart model choice is the bigger lever.

The pricing page also notes that Claude 4.7 and later models use a tokenizer that "produces approximately 30% more tokens for the same text." Every model on the chart is post-4.7, so the comparisons between them are consistent.

## Bottom line

The 3x figure is accurate for Claude Fable 5.1 going from low to max effort on Terminal-Bench 3.0. Hardware, not security, gained the most from higher effort. The benchmark covers terminal coding tasks only, so applying it to finance or compliance is an analogy. Switching from Fable 5.1 to Opus 5.5 can cut cost more than lowering effort.

## Sources

- [Spending your effort, claude.dev](https://claude.dev/blog/spending-your-effort/) - the chart, token medians, domain table and effort guidance
- [claude.dev](https://claude.dev/) - the blog's description
- [Thariq Shihipar, ai.engineer](https://ai.engineer/speakers/thariq-shihipar) - author's role at Anthropic
- [Anthropic pricing](https://platform.claude.com/docs/en/about-claude/pricing) - model prices and tokenizer note
- [Introducing Claude Opus 5.5, Anthropic](https://www.anthropic.com/claude-opus-5-5) - effort levels and cost comparison
- [Terminal-Bench 4.0 leaderboard writeup, Snorkel AI](https://snorkel.ai/leaderboard/terminal-bench-4-0/) - Terminal-Bench 3.0 task count
- [Terminal-Bench](https://www.tbench.ai/) - the benchmark's site
- [@forgedlabsai on TikTok](https://www.tiktok.com/@forgedlabsai) - the creator's profile


## Video description and verification notes

**The file is a 67.93-second MP4 at 720x1280, 30 fps measured with ffprobe (the `fps` field in the yt-dlp metadata is null), HEVC video and AAC stereo audio at 44.1 kHz, overall container bitrate 189,383 bps against a reported `tbr` of 189, audio stream 48,115 bps, 1,608,185 bytes on disk (1.53 MiB), SDR, carrying the format id `bytevc1_720p_189382-1 - 720x1280`. It posted 2026-10-05T22:43:26 UTC from the handle @forgedlabsai, whose channel nickname is Forged Labs, over an audio track listed as "original sound" and credited to the artist "Forged Labs." At the 2026-10-07 capture, two calendar days and roughly 25 hours after posting, the video showed 369 views, 13 likes, 0 comments, 3 reposts, and 6 saves. I read all 34 frames in `tokdoc_frames/` and all 130 words of `transcript.txt`. Every frame carries the same composition: a full-bleed dark-mode screenshot of a line chart with a webcam picture-in-picture pinned to the lower right. The chart header reads "Terminal-Bench 3.0" with the subhead "Pass rate against tokens spent, by effort setting" and a small "PNG" label at the top right. The y-axis is labeled "Share of attempts that passed" with gridlines at 20%, 30%, 40%, 50%, 60%, and 70%; the x-axis label reads "Median tokens pe" before the PiP cuts it off, with visible ticks "50k" and "100". Four series run bottom-left to top-right and are labeled at their right-hand endpoints "Opus 5.5 / max" (solid white, highest), "Fable 5.1 / max" (salmon), "Opus 5 / max" (dashed grey), and "Fable 5 / max" (dotted grey). Individual points carry the effort labels "low", "medium", "high", "xhigh", and "max". A two-line annotation sits near the 60% gridline reading "same score, / half the tokens", positioned beside the Opus 5.5 "high" point and level with the Fable 5.1 "max" point. Frames 001 and 002 alone add a white rounded card over the chart's upper third with purple text reading "Don't pay AI to / Overthink Everything"; the card is gone by frame 003 and never returns. Burned-in captions advance through the narration with single keywords highlighted in yellow: "big money" (frames 001-002), "anthropic" (005-007), "three times" (008-010), and "useful" (029). The speaker is a bald man with short grey stubble, a grey crewneck tee, a lavalier capsule visible at his collar in frames 029-030, an in-ear bud, and a forearm tattoo visible in frames 008 and 033, lit by a soft frontal key against a plain interior. The caption text across the full set reads, in order: "most companies are about to waste big money" (001-002), "by asking AI to think deeply about everything" (003-004), "an anthropic engineer just tested different AI effort levels" (005-007), "and found that higher effort used roughly three times as many tokens" (008-010), "it substantially improved work with the hidden edge cases" (011-012), "especially security" (013), "but delivered much smaller gains on some routine tasks" (014-015), "the business lesson is simple" (016), "match AI effort to the cost of being wrong" (017-018), "use fast inexpensive settings for brainstorming" (019-021), "drafts and early iterations" (022), "and increase the effort for financial analysis" (023-024), "security reviews customer commitments" (025), "compliance and final verification" (026-027), "my verdict this is useful now" (028-029), "don't give every task the same AI budget" (029-030), "spend more reasoning where errors create real consequences" (031-032), and "not where an employee can quickly review the result" (033-034).**

## The chart on screen is Anthropic's, the author is an Anthropic engineer, and the three-times-tokens figure is accurate

The narration says, "An anthropic engineer just tested different AI effort levels and found that higher effort used roughly three times as many tokens in one benchmark." Both halves check out.

The chart in every frame is reproduced from [Spending your effort](https://claude.dev/blog/spending-your-effort/), published on claude.dev on September 25, 2026 and bylined Thariq Shihipar. The site's own header describes itself as a "Blog / technical writing for people building with Claude" and the tagline reads "Sharing tips, tricks, and POVs from Anthropic's developers," which I confirmed by fetching [claude.dev](https://claude.dev/) on 2026-10-07. Shihipar's speaker bio on [ai.engineer](https://ai.engineer/speakers/thariq-shihipar), fetched the same day, states he is a "member of technical staff at Anthropic working on Claude Code." The video's "anthropic engineer" description is correct, and the chart's "Terminal-Bench 3.0" header matches the article's figure exactly, including the "same score, half the tokens" annotation visible in frames 003 through 034.

The token multiple comes from the article's run counts. It reports Claude Fable 5.1 at low effort as "370 attempts, median 73k tokens each" and at max effort as "370 attempts, median 222k tokens each." That is 3.04 times the median token spend. "Roughly three times" is right to two significant figures. The 370 attempts also reconcile cleanly with the benchmark's size: Terminal-Bench 3.0 shipped in July 2026 with 74 tasks, per the [Snorkel AI leaderboard writeup](https://snorkel.ai/leaderboard/terminal-bench-4-0/), and 74 tasks at five attempts each is 370.

What the video drops is which model the figure describes. The 73k-to-222k sweep is Claude Fable 5.1, not any Opus model, and it is low effort against max effort rather than a general statement about "higher effort." The chart on screen plots four separate models with four separate curves, and the narration never names any of them.

## Hardware, not security, posted the largest gain in the article's own domain table

Frame 013 pairs the caption "especially security" with the chart still on screen. The article does break out pass rates by domain at low effort versus max effort, and the numbers are:

| Domain | Low | Max | Change |
|---|---|---|---|
| Security | 64% | 87% | +23 points |
| Hardware | 34% | 75% | +41 points |
| ML | 54% | 73% | +19 points |
| Science | 41% | 61% | +20 points |
| Software | 43% | 56% | +13 points |
| Media | 18% | 30% | +12 points |
| Operations | 12% | 22% | +10 points |

Security improved, so the direction of the video's claim is right. The superlative is wrong. Hardware gained 41 percentage points, nearly double security's 23, and hardware more than doubled its pass rate while security rose by about a third of its starting value. If the video wanted one headline domain, the data points at hardware.

Where the security framing does hold is in the article's worked example. It names `html-js-filter`, "a Terminal-Bench 3.0 task that asks for an HTML sanitizer that strips every way of smuggling JavaScript into a page," and reports that "Fable 5.1 went from 1/5 at low to 5/5 at xhigh." That is a security-flavored task with a clean effort response, and it is likely what the video is compressing. One task is not the same thing as a domain ranking.

## "Much smaller gains on some routine tasks" is a fair read of the absolute numbers and a misleading read of the relative ones

The transcript says higher effort "delivered much smaller gains on some routine tasks." Operations and media are the two smallest absolute movers in the table, at +10 and +12 points. On that measure the claim stands.

Two things complicate it. First, the article never calls operations or media "routine." They are the two domains where both models score worst in absolute terms, which is a different problem from the task being easy. Operations tops out at 22% even at max effort. Second, in relative terms operations rose from 12% to 22%, which is an 83% increase in pass rate, and media rose from 18% to 30%, a 67% increase. Those are larger relative movements than security's 64% to 87%. The video picks the framing that supports its thesis without saying which framing it picked.

## The chart's loudest claim is one the narration never makes

The annotation sitting at the 60% gridline in frames 003 through 034 reads "same score, half the tokens." On the chart it is placed beside the Opus 5.5 "high" point and level with the Fable 5.1 "max" point, which sits at roughly twice the x-position. The article frames this as the generational improvement: "Fable 5.1 and Opus 5.5's effort curves are our best yet: at each level, there is an uptick in benchmark scores and tokens consumed."

That is a claim about picking a better model, and it is independent of the video's thesis about picking a lower effort level. Vendor pricing makes the gap wider than the token count alone. Per [Anthropic's pricing page](https://platform.claude.com/docs/en/about-claude/pricing), fetched 2026-10-07, Claude Opus 5.5 is $4/MTok input and $20/MTok output, while Claude Fable 5.1 is $10/MTok input and $50/MTok output. Opus 5.5 costs 40% of Fable 5.1 per token on both sides. [Anthropic's Opus 5.5 launch page](https://www.anthropic.com/claude-opus-5-5), fetched the same day, states that "Opus 5.5 at default effort beats Opus 5 at max effort for about a fifth of the cost." A viewer who acts only on the video's advice tunes the effort dial and leaves the model choice untouched, which is the larger of the two levers on this chart.

One caveat on reading any token count across model generations: the same pricing page notes that "Claude 4.7 and later models and Claude Mythos Preview use a newer tokenizer" that "produces approximately 30% more tokens for the same text." Every model on this chart is post-4.7, so the comparison between them is internally consistent.

## The source tested 74 terminal coding tasks; the video extends it to compliance and customer commitments

Frames 023 through 027 carry the operational advice: "increase the effort for financial analysis," "security reviews customer commitments," "compliance and final verification." The posted description repeats the list, naming "security, financial decisions, compliance, and final verification."

None of that is in the source. The article's domains are hardware, science, machine learning, software engineering, security, operations, and media, all measured by whether an agent completes a terminal task. [Terminal-Bench](https://www.tbench.ai/) is described by its maintainers as "a benchmark to measure and evolve with the frontier of agent work," hosted by Stanford, Harbor, and the Laude Institute, and the Snorkel writeup describes it as evaluating whether agents can operate real terminals, "writing and debugging code, configuring environments, and recovering from failure, all without a GUI." There is no financial analysis task, no compliance task, and no customer-commitment task in it.

The article's own guidance is written for coding and stays there. Low is "for when I want quick responses that are in the loop, e.g. brainstorming, sketching, easy changes." Medium is "for most of my regular software engineering work, e.g. new feature implementation." High is "for work where verification is important or there are edge cases, e.g. fixing a bug in a brownfield codebase." Max is "When I want Claude to operate fully autonomously to solve difficult problems, e.g. end to end building and verification of an app."

The video's extension to business workflows is a reasonable analogy. It is presented as a finding. The opening line, "Most companies are about to waste big money by asking AI to think deeply about everything," attaches no spend figure, no company count, and no source, and nothing in the article supports it.

## This post is not selling anything I could find

The description names no product, carries no link, and ends with hashtags only: #ArtificialIntelligence #AICosts #BusinessStrategy #AILeadership #Productivity. The transcript contains no call to action. The handle @forgedlabsai is unrelated to the Minecraft creator "Forge Labs" (@forgelaboratories) that dominates search results for that name. I fetched [the uploader's profile page](https://www.tiktok.com/@forgedlabsai) on 2026-10-07 and it returned only TikTok's generic "Make Your Day" shell with no bio, link, or follower count, so I cannot confirm what, if anything, the account promotes elsewhere. On the evidence in this video, it is commentary rather than an ad.

## Key Takeaways

- **Verified:** The "roughly three times as many tokens" figure is accurate. Anthropic's article reports Claude Fable 5.1 on Terminal-Bench 3.0 at a median 73k tokens per attempt at low effort and 222k at max effort across 370 attempts each, a ratio of 3.04.
- **Verified:** The "anthropic engineer" attribution is correct. Thariq Shihipar is listed as a member of technical staff at Anthropic working on Claude Code, and claude.dev describes itself as carrying POVs from Anthropic's developers.
- **Verified:** Higher effort did substantially improve the security domain, from 64% to 87% pass rate, and the article's `html-js-filter` HTML sanitizer task went from 1/5 at low effort to 5/5 at xhigh.
- **Correction:** Security was not the standout domain. Hardware gained 41 percentage points (34% to 75%) against security's 23. The video's "especially security" picks the second-largest gain.
- **Correction:** The video presents the 3x figure as a property of "higher effort" in general. It is one model (Claude Fable 5.1) at the two extremes of a five-level scale (low versus max) on one benchmark. The chart on screen plots four different models with four different curves and the narration names none of them.
- **Partial correction:** "Much smaller gains on some routine tasks" is true in percentage points for operations (+10) and media (+12), and false in relative terms, where operations rose 83% and media 67% from their starting pass rates. The article does not label either domain routine.
- **Partial correction:** The advice list in frames 023-027 (financial analysis, customer commitments, compliance) describes work the source never tested. Terminal-Bench 3.0 is 74 terminal coding tasks. The extrapolation may be sound; it is not a finding from this data.
- **Unverified:** "Most companies are about to waste big money." No spend figure, no company count, and no source appear in the video or in the article it draws from.
- **Context:** The chart's own annotation, "same score, half the tokens," is about choosing Opus 5.5 over Fable 5.1, not about choosing a lower effort level. Opus 5.5 is also priced at 40% of Fable 5.1 per token on both input ($4 versus $10 per MTok) and output ($20 versus $50 per MTok).
- **Context:** Anthropic's pricing page notes that Claude 4.7 and later use a tokenizer producing about 30% more tokens for the same text. Every model on this chart is post-4.7, so the token comparisons between them are consistent.
- **Context:** The video names no product and carries no link. The uploader's profile page would not load any bio or link for me on 2026-10-07, so I cannot rule out promotion elsewhere on the account.

## Resources

- [Spending your effort, Thariq Shihipar, claude.dev, September 25, 2026](https://claude.dev/blog/spending-your-effort/): the source of the on-screen chart, the 73k/222k token medians, the 370-attempt counts, the per-domain pass-rate table, the `html-js-filter` example, and the low/medium/high/max guidance.
- [claude.dev](https://claude.dev/): establishes the blog's own description as "technical writing for people building with Claude" sharing "POVs from Anthropic's developers."
- [Thariq Shihipar speaker bio, ai.engineer](https://ai.engineer/speakers/thariq-shihipar): confirms he is a member of technical staff at Anthropic working on Claude Code.
- [Anthropic pricing documentation](https://platform.claude.com/docs/en/about-claude/pricing): the live per-million-token prices for Claude Opus 5.5 ($4/$20), Claude Opus 5 ($5/$25), Claude Fable 5.1 ($10/$50), and Claude Fable 5 ($10/$50), plus the note on the newer tokenizer producing about 30% more tokens.
- [Introducing Claude Opus 5.5, Anthropic](https://www.anthropic.com/claude-opus-5-5): confirms the model name, the low/medium/high/xhigh effort levels with medium as default, and the statement that Opus 5.5 at default effort beats Opus 5 at max effort for about a fifth of the cost.
- [Terminal-Bench 4.0 leaderboard writeup, Snorkel AI](https://snorkel.ai/leaderboard/terminal-bench-4-0/): establishes that Terminal-Bench 3.0 shipped in July 2026 with 74 tasks and that the benchmark measures agents operating real terminals without a GUI.
- [Terminal-Bench](https://www.tbench.ai/): the benchmark's own site, confirming it is hosted by Stanford, Harbor, and the Laude Institute.
- [@forgedlabsai on TikTok](https://www.tiktok.com/@forgedlabsai): the uploader's profile; it loaded but returned no bio, link, or follower data, which is why the promotion question stays open.
