Opus 5 has five effort levels. Higher effort takes longer and consumes limits faster. On a well-specified front-end task, Low effort hit every correctness marker Max did while running at least 7 times faster. Effort is a spend dial, not a quality dial, and the vendor says so in the UI.
- Five effort levels: Low, Medium, High (default), Extra, Max. The API names are
low,medium,high,xhigh,max. - On a single, tightly-specified task, Low matched Max on every correctness marker. Max took at least 7 times longer and produced 80% more code.
- Opus 5 refused a fabricated API parameter and cited its source. It computed a WCAG contrast ratio correctly to two decimals, beating the author's hand arithmetic.
- Thinking is on by default. The
budget_tokensparameter is removed. The minimum cacheable prompt length dropped from 1,024 to 512 tokens. - Two new betas: mid-conversation tool changes and automatic model fallback for flagged requests.
- This is n=1 per condition. One task, one run each. The class says what it does not know.
Section 01
What Opus 5 Is
Anthropic released Claude Opus 5 on July 24, 2026. Model string: claude-opus-5. Context window: 1,000,000 tokens. Max output: 128,000 tokens. Thinking is enabled by default.
It sits between Fable 5 (the most capable Claude, $10/$50 per million tokens) and Sonnet 5 ($3/$15). Opus 5 matches its predecessor Opus 4.8 at $5 input / $25 output per million tokens. Same price, more capable.
Available on the Claude API, claude.ai (default on Max, strongest on Pro), Amazon Bedrock, Microsoft Foundry, Snowflake Cortex AI, GitHub Copilot, and Claude Code.
Anthropic's published benchmarks
These are the vendor's numbers. I am citing them because they exist, not because I ran them. Every figure here is graded ASSERTED (Anthropic).
| Benchmark | Opus 5 Result | Context |
|---|---|---|
| Frontier-Bench v0.1 | 43.3% | More than 2x Opus 4.8 (18.7%) |
| CursorBench 3.2 | Within 0.5% of Fable 5 | At max effort, half the cost |
| ARC-AGI 3 | 30.2% | ~3x next-best model (High effort) |
| OSWorld 2.0 | 70.6% | Beats Fable 5 at ~1/3 the cost |
| AA Intelligence Index | Score 61 | Ranked #1 at launch (Max reasoning) |
Anthropic also disclosed that Opus 5 "remains behind Mythos 5 on cybersecurity tasks." I include that because a vendor volunteering a weakness is more useful than one hiding it.
Source: anthropic.com/news/claude-opus-5, fetched 2026-08-01.
Section 02
The Effort Dial
That tooltip is the honest frame for this entire class. The vendor is telling you effort is a spend dial. Not a quality dial. Not an intelligence dial. A spend dial.
The API side
The API parameter is output_config: {"effort": "high"}. Five values: low, medium, high (default), xhigh, max. The UI's "Extra" maps to the API's xhigh.
The old budget_tokens parameter for extended thinking is gone. Sending it to Opus 5 returns a 400 error. Thinking is now on by default, and the model decides how deeply to think based on task complexity. You control the ceiling with effort, not the floor.
You can disable thinking only at high or below. Attempting thinking: {"type": "disabled"} with xhigh or max also returns 400.
Source: Anthropic docs, models page, fetched 2026-08-01. Grade: ASSERTED (Anthropic).
Section 03
The A/B Test
Identical prompt, two fresh chats, one variable. The task: a Liquid snippet rendering four metric tiles from a ~~~/| delimited string, with CSS grid, wrapping at 390px, scoped CSS. Run in my Claude Max account, 2026-08-01.
| Low Effort | Max Effort | |
|---|---|---|
| Wall-clock time | 21 seconds or less | 150 to 220 seconds |
| Code lines | 77 | 139 |
| Code characters | 2,215 | 3,991 |
Correct ~~~ split | Yes | Yes |
Correct | split | Yes | Yes |
| CSS grid | Yes | Yes |
| 390px breakpoint | Yes | Yes |
escape filter (XSS) | Yes | Yes |
| Usage comment header | Yes | Yes |
role="list" a11y | Yes | Yes |
Max effort took at least 7 times longer and produced 80% more code. It hit the identical correctness markers. Every requirement the prompt specified was satisfied by both runs. Both independently added output escaping and a role="list" accessibility attribute that the prompt never asked for.
On this task, the extra 62 lines from Max are more configuration surface, not more correctness.
Timing honesty
My polling granularity was 10 seconds, so these are bounds, not stopwatch readings. The ratio is at least 7x; it could be more. I am not going to sharpen a number I did not measure precisely.
The near-miss
I initially wrote that Max had added output escaping that Low lacked. I went back and checked the Low run before publishing. Low had it too. The claim would have been wrong and it would have been flattering. This is exactly the failure mode this Academy exists to teach.
Grade: MEASURED, n=1 per condition, single task. A reader who concludes "Low effort is always enough" has over-read this result, and I am saying so now.
Section 04
The Hallucination Trap
The prompt was a trap. It asked what the allowUnsafeOverwrite argument does on Shopify's themeFilesUpsert GraphQL mutation and what its default is. That argument does not exist.
Opus 5's first words: "No such argument exists."
It then did more than refuse. It enumerated the two real arguments, named the input object and its two fields (body, filename), stated the reference page it fetched and the date it fetched it, and graded its own answer MEASURED.
A model that invents a plausible API parameter tells you something about every other answer it gave that day. This one refused the premise and showed its source.
Grade: MEASURED. Passed.
Section 05
Exact Arithmetic
Foreground #A9BBAE on background #081710. Real DDS palette colors. I asked for a WCAG 2.x contrast ratio.
| Source | Ratio |
|---|---|
| Opus 5 | 9.13:1 |
| Independent computation (Python, relative luminance) | 9.1254 rounds to 9.13:1 |
| My hand arithmetic | 9.12:1 (wrong) |
Opus 5 was right. I was wrong. My hand calculation accumulated rounding error through the ln/exp steps and landed a hundredth low. The model matched the exact computation and correctly reported the pair passes AA (4.5:1) and AAA (7:1) for body text.
I am reporting this honestly because the alternative, quietly dropping the row where the model beat me, is the thing this Academy tells people not to do.
Grade: MEASURED and independently verified.
Section 06
The Two Betas
Both shipped alongside Opus 5 on July 24, 2026.
Mid-conversation tool changes
Developers can now add or remove tools between conversation turns without resending the entire tool list or invalidating the prompt cache. Previously, any change to the tool list broke the cache and forced a full re-send.
Activation requires the beta header mid-conversation-tool-changes-2026-07-01.
For agentic coding, this matters because a long-running agent session often starts with one set of tools (file operations) and later needs a different set (deployment commands). Without this beta, switching tools mid-conversation flushes the cache.
Automatic model fallback
When a request triggers safety classifiers, the API can now auto-route it to a recommended fallback model instead of blocking the response entirely. A "default" fallbacks mode handles model selection. Developers do not need to maintain their own fallback list.
Activation requires the beta header server-side-fallback-2026-07-01.
This converts a hard block into a graceful downgrade. The agent session stays alive instead of stopping on a false positive.
Source: Anthropic docs, fetched 2026-08-01. Grade: ASSERTED (Anthropic).
Section 07
What Changed for Agentic Coding
Four things changed from Opus 4.8 to Opus 5 that matter if you build with AI agents.
- Thinking is on by default. Opus 4.8 required explicit configuration. Opus 5 thinks by default and decides how deeply to reason based on the task. You control the ceiling with effort, not the floor.
- Native self-verification. The model checks its own work without being told. Anthropic advises removing "include a final verification step" from system prompts. The model does it natively.
- Prompt caching is cheaper. Minimum cacheable prompt length dropped from 1,024 tokens to 512 tokens. Agent sessions that re-use the same system prompt across many turns pay less.
- More autonomous, less conservative. Opus 4.8 often asked for clarification. Opus 5 is more likely to proceed with what it infers. On underspecified tasks this means deeper exploration. On well-specified tasks, it means faster completion.
User feedback indicates existing "step by step" or "verify before proceeding" prompt patterns that worked for 4.8 can cause over-verification in Opus 5. Strip them and let the model's native reasoning handle it.
Sources: Anthropic blog and docs, fetched 2026-08-01. Grade: ASSERTED (Anthropic).
Section 08
Safety Posture
Anthropic states Opus 5's safety classifiers trigger approximately 85% less frequently than Fable 5's. The number appeared in press coverage citing Anthropic estimates. The exact methodology is not published. Treat it as directional, not precise.
What it still declines: binary vulnerability scanning, penetration-testing code generation, and explicit exploit construction. Anthropic frames this as a boundary across the model lineup, not a per-tier setting.
What this means in practice: the model is more permissive for legitimate development tasks (deployment scripts, security configurations, authentication flows) while maintaining the same refusal boundary on offensive security.
The automatic model fallback beta (Section 06) is the companion feature. If a request does trigger a classifier, the system can route to a fallback model rather than stopping the session.
Grade: ASSERTED (Anthropic). The 85% figure is directional.
Section 09
What This Class Does Not Know
The graded ledger is the differentiator. A masterclass that only shows wins is not this Academy.
| Gap | Status |
|---|---|
| Head-to-head vs Fable 5 or Opus 4.8 | Not run |
| Underspecified or multi-step tasks where effort may matter more | Not tested |
| Per-task token counts | Unmeasured (UI does not expose them) |
| Cost-per-task comparison | Cannot compute without token counts |
| Timing precision | Coarse (10 s polling). Bounds, not stopwatch readings |
| Harness prompts 04, 05, 11 | Not run. Would strengthen the class but are not load-bearing |
| Benchmark independent replication | None found for any of Anthropic's published figures |
| 85% fewer classifier triggers | Anthropic estimate, methodology not published |
"Extra" = API xhigh mapping | Inferred from matching 5 UI / 5 API levels, not explicitly documented |
This is n=1 per condition. One task, one run each. It is a demonstration, not a benchmark. A reader who concludes "Low effort is always enough" has over-read it.
Bottom Line
Opus 5 ships five effort levels and the vendor tells you in the UI that higher effort consumes limits faster. On a well-specified front-end task, Low matched Max on every correctness marker while running at least 7 times faster. The model refused a fabricated API parameter, cited its source, and computed a WCAG contrast ratio more accurately than I did by hand.
That does not mean Low is always enough. It means on a task where the requirements are clear and the scope is small, the cheapest setting did the job. Where effort matters most, Anthropic's own benchmarks suggest, is on underspecified, multi-step, exploratory work. This class did not test that. Saying so is the point.
FAQ
Frequently Asked Questions
What are the five effort levels in Claude Opus 5?
Low, Medium, High (default), Extra, and Max. In the API they map to low, medium, high, xhigh, and max. Higher effort means more thorough reasoning but consumes your usage limits faster. Anthropic states this directly in the UI tooltip.
Does higher effort mean better code?
Not necessarily. In the measured A/B in this class, Low effort and Max effort hit identical correctness markers on a well-specified Liquid task. Max took at least 7 times longer and produced 80 percent more code, but every requirement the prompt specified was satisfied by both runs. On underspecified or multi-step tasks, higher effort may produce meaningfully different results. This class tested one task at n equals 1 per condition.
How much does Claude Opus 5 cost?
5 dollars per million input tokens and 25 dollars per million output tokens. This matches Opus 4.8 exactly and is half the price of Fable 5 at 10 and 50 dollars respectively. These are Anthropic published prices as of July 24, 2026.
What is the effort dial in the API?
The API uses output_config with an effort key. Valid values are low, medium, high, xhigh, and max. The default is high. The previous budget_tokens parameter for extended thinking has been removed. Sending budget_tokens to Opus 5 returns a 400 error.
Can Opus 5 hallucinate API parameters?
In the measured test in this class, Opus 5 was asked about a fabricated argument on a real Shopify mutation. It responded with No such argument exists, enumerated the real arguments, named its source, and graded its own answer. A model that invents plausible parameters tells you something about every other answer it gave that day.
What changed from Opus 4.8 to Opus 5?
Thinking is enabled by default instead of requiring configuration. The model self-verifies natively so you can remove verification step instructions from prompts. The minimum cacheable prompt length dropped from 1024 to 512 tokens. Two new betas shipped: mid-conversation tool changes and automatic model fallback for flagged requests.
Is Opus 5 less restrictive than Fable 5?
Anthropic states Opus 5 safety classifiers trigger approximately 85 percent less frequently than Fable 5. It still declines exploit generation, binary vulnerability scanning, and penetration testing code. The automatic model fallback beta can route flagged requests to a fallback model instead of blocking them entirely.
What benchmarks did Anthropic publish for Opus 5?
Frontier-Bench v0.1 at 43.3 percent, CursorBench 3.2 within 0.5 percent of Fable 5 at max effort, ARC-AGI 3 at 30.2 percent, OSWorld 2.0 at 70.6 percent, and AA Intelligence Index score of 61 at number 1 ranking. All are vendor-reported figures marked ASSERTED in this class. No independent replication was found for any of them.
What does this masterclass not test?
It does not test Opus 5 against Fable 5 or Opus 4.8 head to head. It does not test underspecified multi-step tasks where higher effort may matter more. It has no token counts because the UI does not expose them. Timing is coarse at 10 second polling. Three harness prompts remain unrun. The class says so explicitly.
What does this masterclass cost?
Nothing. This masterclass is free. No signup, no email, no paywall, no certificate requirement. It is part of the DDS Vibe Academy, a free AI coding curriculum built by Robert McCullock at Design Delight Studio.