Claude Opus 5 makes rework the real model price
Two models can produce acceptable first drafts. One costs more to run. The other sends a senior employee back into the document three times. Only one of those costs appears neatly on the invoice.
Claude Opus 5 is interesting because Anthropic is positioning it between everyday models and its highest-end Fable 5: close to frontier performance at half the price, with stronger self-checking and longer-horizon work.
That does not make it the default choice. It makes it a candidate for work where correction time is already expensive.
- The Big Signal: Price Opus 5 against correction time, not easy prompts
- Build This: Calculate when a premium model pays for itself
- Worth Watching: AI is already pulling work across job boundaries
Claude Opus 5 moves the comparison from token cost to completion cost
Use it for routine work and you may simply spend more. Use it where a cheaper model creates repeated repair and the economics can reverse.
Anthropic released Claude Opus 5 on 24 July. The company says it approaches the intelligence of Claude Fable 5 at half the price and improves substantially on Opus 4.8 without increasing the base API rate.
- Access: Available across Claude’s supported platforms, the default model on Claude Max and the strongest model offered through Claude Pro.
- API price: $5 per million input tokens and $25 per million output tokens, unchanged from Opus 4.8.
- Fast mode: Approximately 2.5 times the normal speed for twice the base price.
- Where Anthropic sees the gain: Difficult coding, analysis, automation, document work and longer tasks that benefit from checking and revision.
- Limitations: Much of the launch evidence comes from Anthropic and early-access partners, while some restricted cyber requests may be blocked or routed to Opus 4.8.
The invoice is only half the cost
Imagine a cheaper model produces a draft in five minutes, then a manager spends 35 minutes finding omissions, checking calculations and rebuilding the recommendation. A more expensive model can still be the lower-cost option if it removes enough of that repair work.
The reverse is also true. A long task may remain easy to verify, while a short legal or financial decision may still need qualified judgement. The useful comparison is a real deliverable with a visible correction history—not a polished demo prompt.
The cheapest model is not the one with the lowest token price. It is the one that reaches a usable answer with the least paid human repair.
the work is routine, low-risk and quick to verify.
Test Opus 5 whenthe task is vague, multi-step or currently comes back for repeated correction.
Keep a qualified person in charge whenthe decision is regulated, high-stakes or cannot be checked against clear evidence.
Run the repair test
Choose one task that was returned for correction at least twice last month. Give the same brief and source material to Opus 5 and your current model. Compare the completed work, correction minutes, checking time and total model cost. Upgrade only when the reduction in repair is worth more than the extra spend.
Read the Claude Opus 5 announcementCalculate your premium-model break-even point
The outcome is a one-page comparison showing whether higher model spend saves enough review time to pay for itself.
You’ll need
- Opus 5 and your current model
- One task with previous corrections
- A spreadsheet and a human reviewer
Five steps
- Select a repaired task. Use a real but reversible deliverable, such as a management memo, spreadsheet analysis, code fix or first-pass contract review.
- Freeze the brief. Give both models the same inputs, tools, output format and success criteria. Do not rescue one version with extra context halfway through.
- Count the repair. Record factual corrections, missing requirements, rewrites, reviewer minutes and whether the task had to be sent back for another pass.
- Price the finished result. Convert model spend into the same currency as labour, then add reviewer cost: hourly rate × correction minutes ÷ 60.
- Write the routing rule. State which tasks remain on the cheaper model, which move to Opus 5 and which still require a specialist from the beginning.
Task tested: [ ]
Required deliverable: [ ]
Reviewer hourly cost: £[ ]
Current-model spend: [ ]
Opus 5 spend: [ ]
Current-model correction minutes: [ ]
Opus 5 correction minutes: [ ]
Current-model total completion cost: [ ]
Opus 5 total completion cost: [ ]
Factual errors after review: [ ] / [ ]
Requirements missed: [ ] / [ ]
Decision: Cheaper model / Opus 5 / Specialist
Routing rule: [ ]
Review date: [ ]
Pro tip
Count final checking time for both models. A stronger model may reduce correction without removing review, and a confident answer is not evidence that the work is correct.
Read Anthropic’s Opus 5 prompting guideAI is absorbing the handoffs around a job
The useful signal is not that everyone becomes a specialist. It is that more work can be attempted before a specialist is called.
OpenAI analysed more than 800,000 messages from US ChatGPT users and found that 43.5% of occupation-specific messages involved tasks associated with another occupation. The share was particularly high among customer-experience workers, designers, HR teams, legal workers and marketers.
Usage does not prove that every task was completed well, and crossing a job boundary does not remove accountability. It does reveal where queues may shrink: drafting, troubleshooting, preliminary analysis and research that previously waited for another department or supplier.
Improve the handoff before trying to eliminate it.
For a small business, the practical move is to define a better first pass. Let AI produce the draft or analysis, require evidence, then send a cleaner package to the person who owns the judgement.
Read the task-crossover researchAxios says it embedded AI-enablement leads across editorial, events and revenue teams, including employees without traditional technical backgrounds. The useful small-business version is narrower: give one person ownership of one workflow, one reviewer and one measurable result—not a vague company-wide transformation brief.
Read the Axios case studyNVIDIA and more than 40 founding partners have launched the Open Secure AI Alliance to develop open models, harnesses and security tools. NVIDIA’s NOOA research framework is already available for testing, tracing, auditing and governing agent behaviour; the wider alliance is still at the building stage.
Explore the Open Secure AI AllianceReply with the task and where the output usually breaks. I am interested in the work that looks cheap until someone has to repair it.
I would start with the task that already comes back twice. That is where a better model has a fair chance to earn its keep.
Until next time,
Sandeep
Future Relay
