Claude Opus 5 for Business: Which Workflows Are Worth Revisiting?
Quick answer
Opus 5 makes complex workflows worth another look when older models lost context, missed steps, or required too much fixing.
Anthropic positions Opus 5 near Fable 5 on many tasks at half Fable's published token price. The business opportunity is not cheaper writing. It is a lower cost of testing multi-step work that failed too often before.
Start with one valuable workflow. Give Opus 5 and your current model the same inputs, tools, success rules, and human reviewer. Compare finish rate, fixing time, review time, and total cost. Keep the model with the lowest cost per accepted result.
Important: Benchmarks show where a model may be worth testing. They do not prove that a workflow is ready to run alone. Data access, human review, legal duties, and sector rules still apply.
What matters for business teams
Opus 5 lowers the cost of testing harder workflows, but the right model still depends on the work, risk, and review needed.
- Anthropic released Claude Opus 5 on July 24, 2026, at $5 per million input tokens and $25 per million output tokens.
- Those published rates are half Fable 5's rates and unchanged from Opus 4.8.
- Opus 5's strongest reported gains appear in multi-step business work, tool use, computer use, web research, automation, and coding.
- Fable 5 retains narrow leads on several coding and legal tests. Mythos 5 retains stronger restricted skills in exploit development and advanced biology research.
- Half the token price does not mean half the workflow cost. Effort settings, retries, tool calls, output length, fixing time, and review time all matter.
- The practical decision is whether a workflow that failed before can now produce useful work often enough to repeat.
Contents
This guide explains what changed, which workflows deserve another test, where stronger models still matter, and how to compare results.
- What changed with Claude Opus 5
- Why half the token price is not half the workflow cost
- Which failed workflows are worth revisiting
- The Attainment Workflow Revisit Test
- Workflow ideas for Attainment buyer segments
- Where Fable 5 and Mythos 5 still differ
- How to compare Opus 5 with your current model
- When not to use Opus 5
- Frequently asked questions
- What to do next
What changed with Claude Opus 5?
Opus 5 brings results near Fable 5 to the same token price as Opus 4.8, making harder recurring workflows cheaper to test.
Anthropic released Claude Opus 5 on July 24, 2026. The company describes it as a model for complex agentic coding and enterprise work that approaches Fable 5 on many tasks while charging half as much per token.
The published Claude API pricing is:
| Model | Input per 1M tokens | Output per 1M tokens | Relative token price |
|---|---|---|---|
| Claude Opus 5 | $5 | $25 | 1x |
| Claude Fable 5 | $10 | $50 | 2x |
| Claude Opus 4.8 | $5 | $25 | 1x |
That pricing changes the decision for teams that found Fable 5 capable but too costly to use often. The same budget buys twice as many Opus 5 tokens when the input and output mix stays the same.
The more important change is what the model can do at that price. Published tests show large gains over Opus 4.8 in computer use, web research, business work, automation, and coding. These skills matter when AI must track a long task, use tools, compare evidence, and finish the work.
Opus 5 also supports a one-million-token context window and can be eligible for Anthropic's API zero-data-retention arrangement. Fable 5 and Mythos 5 require 30-day retention. Platform, feature, and contract terms still need separate review, but data retention can matter more than model price for some teams.
Why is half the token price not half the workflow cost?
Published token prices fall by half versus Fable 5, but the useful measure is total cost for an output your team accepts.
Token price is one line in the cost equation. A workflow also uses tool calls, retries, cached and uncached context, model effort, setup time, fixing time, and expert review.
A lower-priced model can cost more if it needs repeated runs or heavy fixing. A higher-priced model can cost less if it finishes correctly on the first attempt. The result can also change when the same model runs at a different effort setting.
Artificial Analysis tested Opus 5 on AA-Briefcase, a test of long business tasks. Opus 5 at high effort scored 1,606 Elo and cost $10.41 per task. Fable 5 scored 1,574 Elo and cost $22.30 per task.
That is useful evidence, but it remains one test under one setup. It does not mean every business gets the same savings.
The operating measure should be:
Cost per accepted result = model and tool cost + fixing time + review time + failed-run cost
An accepted result is work an expert reviewer would use. Draft volume, token volume, and benchmark scores are extra measures. Accepted work is the business measure.
Which failed workflows are worth revisiting?
Revisit work that failed because older models lost context, missed steps, or required too much expert fixing to justify repeat use.
Basic drafting, summarization, and simple classification were already available with earlier models. Opus 5 matters when the failure happened later in the task.
The strongest workflows have several of these traits:
- Inputs are spread across documents, spreadsheets, inboxes, calls, and business systems.
- The work requires multiple steps and tools, not one response.
- Rules change or conflict as the task moves forward.
- The final work must preserve sources, decisions, open issues, and ownership.
- An expert spent more time fixing the AI than the AI saved.
- The task repeats often enough for a better finish rate to matter.
The model cannot repair a workflow with missing data, unclear ownership, or no definition of success. It can only work with the inputs, permissions, tools, and review process provided.
This distinction protects teams from buying more AI before proving which workflow is worth fixing.
What is the Attainment Workflow Revisit Test?
Choose repeated, valuable work with clear inputs, an accepted finish line, and an expert who can review the completed result.
Use five questions before testing Opus 5:
1. Is the workflow valuable and repeated?
Choose work that happens weekly, monthly, or at a clear trigger. A rare task may be important, but it gives you less proof and fewer chances to improve the process.
2. Did older models fail to finish?
Look for missed rules, lost context, weak tool use, poor source matching, incomplete work, or too much expert fixing. Disliking the writing style is not enough.
3. Can you define accepted output before the run?
Write the finish line first. Name the required sections, sources, math, approvals, open issues, and file format. If success is subjective, two reviewers may judge the same output differently.
4. Can a qualified human review the result?
AI can prepare, compare, draft, flag, and route. An expert should keep control of clinical, legal, buying, compliance, investment, pricing, and external messages.
5. Can you measure the economics?
Record finish rate, fixing time, review time, total model and tool cost, and accepted output. Compare the whole workflow, not one model response.
If all five answers are clear, the workflow is a strong candidate. If two or more are unclear, fix the process before changing the model.
Which Attainment buyer workflows are worth revisiting?
The strongest workflows combine long context, several tools, real business value, and a clear human approval boundary.
The examples below are ideas to test, not claims of proven use. Start with one bounded workflow and non-sensitive test data whenever possible.
Specialty clinics: consult-path exception review
A specialty clinic can test whether Opus 5 can bring inquiry sources, call notes, booking records, and follow-up history into one consult-path action list.
Older models often struggled because the same inquiry appeared under inconsistent labels across phone, form, CRM, and scheduling records. A useful output must show what happened, what is missing, and which item requires staff attention.
The model should not diagnose, triage, recommend treatment, promise compliance, or send patient-facing messages without approval. Staff keep every clinical and relationship decision. Start with the least access needed and measure accepted items, false flags, review time, and the team's view of the work.
PE and search operators: one 100-day workflow review
A PE-backed, search-funded, or permanent-capital operator can test Opus 5 on one operating workflow using management reports, CRM exports, SOPs, dashboards, and interview notes.
The target is a traceable workflow diagnosis, not a broad operating plan. The output should identify the stuck step, source evidence, missing data, current owner, proposed measure, and the smallest proof sprint management could approve.
Management selects the priority and approves any change. The model should not make investment, valuation, staffing, or legal decisions. Measure source accuracy, accepted findings, fixing time, and whether the output creates one clear operating decision.
Funded organizations: evidence-gap and ownership mapping
A team that wins, manages, or delivers funded work can test Opus 5 on one grant, report, renewal, or funded-program workflow.
The model can compare the agreement, report template, prior submissions, delivery records, and internal notes to produce a list of rules, available proof, missing proof, owners, and deadlines.
Program staff confirm fit, compliance, claims, outcomes, and submissions. The model prepares the map and draft support. It does not decide what qualifies or what should be said to a funder. Measure missed rules, claims with no source, review time, and the view of each deadline.
Government contractors: live pursuit rules and ownership
A government contractor can test Opus 5 on one live pursuit using the solicitation, amendments, questions and answers, partner inputs, past-performance records, and internal deadlines.
The target is a rules and ownership map that stays current as documents change. Older models often lost track when an amendment changed a rule or when proof lived outside the main solicitation.
Humans retain bid, pricing, legal, procurement, compliance, and submission decisions. Use limited access and approved material. Measure missed rules, source accuracy, update time, ownership gaps, and reviewer confidence.
Home services: booked-job recovery action list
A home-services operator can test Opus 5 on one branch, service line, or week by bringing together calls, estimates, dispatch records, CRM stages, and follow-up history.
The target is a clean list of calls, estimates, and follow-ups that need human action, with the source and reason attached. This is harder than summarizing calls because the model must determine current status across several systems.
Staff approve customer messages, pricing, scheduling, and service decisions. Measure accepted items, false positives, review time, and the share of records with enough proof to act.
Other Attainment buyer paths
| Buyer path | Candidate workflow | Required human control |
|---|---|---|
| Cybersecurity and IT | Assemble a buyer-ready proof package from technical evidence, prior answers, controls, and expert notes. | A qualified expert approves every security, risk, and compliance statement. |
| B2B SaaS and AI startups | Reconcile customer calls, CRM stages, product documents, website copy, and loss notes into a positioning-to-conversion diagnosis. | Leadership selects positioning, offer, pricing, and market actions. |
| Professional services firms | Build a first proposal or engagement draft from referral context, approved language, prior work, and expert notes. | A qualified professional approves advice, scope, terms, and client communication. |
| Multi-location operators | Compare one workflow across locations and produce a variance map. | Operations leaders decide what to standardize and what stays local. |
| Real estate and construction | Reconcile project documents, change requests, approvals, schedules, and follow-up into an exception list. | Humans approve contracts, changes, schedules, payments, and legal interpretations. |
| Education and workforce programs | Build an outcome-evidence map from program records, funder rules, delivery notes, and report templates. | Program leaders approve outcomes, fit, and compliance claims. |
| Healthcare vendors | Assemble buyer education or proposal material from approved product proof, prior answers, and buyer needs. | Clinical, regulatory, legal, and procurement reviewers approve their domains. |
These workflows share the same pattern: the model handles prep and comparison while the person who owns the work keeps the decision.
Where do Fable 5 and Mythos 5 still differ?
Fable 5 remains stronger for some highest-difficulty work, while Mythos 5 keeps restricted cyber and biology skills.
Opus 5 and Fable 5 sit in the same top model group, but they are not interchangeable. Anthropic's reported tests show Fable 5 retaining narrow leads on several coding, legal, and no-tool reasoning tasks.
Fable 5 may still justify its higher token price when a small skill edge cuts expert review or failure risk. That decision should come from the workflow test, not the model name.
Fable 5 and Mythos 5 share the same underlying model, with different safeguards and access policies. Fable 5 is generally available and Mythos 5 remains limited.
Mythos 5 remains stronger in turning software flaws into working exploits and in advanced autonomous biology research. Those restricted skills are not ordinary business workflow features and should not be used to describe Opus 5.
For most teams, the practical choice is between Opus 5, Fable 5, a lower-cost model, and the current human process. Mythos belongs in approved programs with expert controls.
How should a business compare Opus 5 with its current model?
Run the same bounded workflow with fixed inputs and acceptance criteria, then compare completed work, review time, and total cost.
Use this six-step test:
- Choose one live workflow. Select repeated work with measurable value and a qualified reviewer.
- Freeze the test inputs. Give each model the same documents, tools, instructions, and access.
- Define accepted output. Write the required sections, sources, decisions, open issues, and file format before either model runs.
- Run more than once. One successful result can be luck. Use enough comparable cases to see a pattern.
- Record the whole cost. Include model charges, tools, retries, fixing time, review time, and failed runs.
- Keep the simplest model that clears the bar. Use Fable 5 only where its extra skill changes the accepted result. Use a lower-cost model where it performs equally well.
Track at least these fields:
| Measure | What it answers |
|---|---|
| Finish rate | Did the model finish every required step? |
| Accepted-result rate | Would the qualified reviewer use the output? |
| Fixing time | How long did the reviewer spend fixing errors? |
| Review time | How long did approval take, even when no fix was needed? |
| Source accuracy | Can each important statement be traced to the supplied evidence? |
| Model and tool cost | What did the run consume? |
| Total cost per accepted result | Did the workflow become affordable to repeat? |
The output is a routing decision, not a universal model ranking. Different steps in the same workflow may need different models.
When should a business avoid using Opus 5?
Do not add a stronger model when the real problem is missing data, unclear ownership, absent controls, or a task that should remain human.
Avoid or delay the workflow when:
- No one can define accepted output.
- The required data is missing, unreliable, or unavailable.
- The workflow owner is unclear.
- A qualified reviewer is unavailable.
- The task makes clinical, legal, compliance, procurement, investment, pricing, employment, or safety decisions without human approval.
- The model would need more system access than the value of the test justifies.
- The task happens too rarely to create reliable evidence or useful payback.
- The current human process changes every time and has not been documented.
A stronger model can reduce execution failures. It cannot decide which business problem matters, create missing evidence, or assume accountability.
Frequently asked questions
These answers cover the main business questions about Opus 5 pricing, skills, model choice, data handling, and safe testing.
Is Claude Opus 5 the same model as Fable 5?
No. Opus 5 is an Opus-class model. Fable 5 and Mythos 5 share a different underlying model. Anthropic says Opus 5 approaches Fable 5 on many tasks, but Fable 5 remains its most capable generally available model overall.
Is Opus 5 the restricted Mythos model released to every business?
No. Mythos 5 remains limited to approved teams. Opus 5 was not trained as a public version of Mythos's restricted offensive-cyber skills.
Does Opus 5 cost half as much as Fable 5?
Its published standard input and output token prices are half Fable 5's. The total workflow saving varies with output length, effort, retries, tools, fixing time, and human review.
What business work is Opus 5 best suited to?
The strongest published evidence supports multi-step business work, coding, computer use, web research, automation, and tasks that require long context. Every live workflow still needs its own test.
Should a business replace its current model with Opus 5?
Only where Opus 5 produces a better accepted result or reduces the total cost of completing the work. Keep lower-cost models for tasks they already complete reliably.
Can a business in a regulated field use Opus 5?
In some cases, with the right contracts, platform controls, minimum access, approved features, and expert human review. Model choice alone does not meet privacy, security, legal, clinical, or industry rules.
Why can data retention affect model choice?
Anthropic's API docs say Opus 5 can be used under some zero-data-retention plans, while Fable 5 and Mythos 5 require 30-day retention. Whether you can use it still depends on the platform, contract, workspace, and API features.
What should a business test first?
Choose one valuable weekly workflow that older models failed to finish reliably. Define accepted output, keep the inputs fixed, run a fair test, and measure cost per accepted result.
What should a business do next?
Opus 5 matters because it may make difficult workflows affordable to repeat, not because every business needs a new model.
The useful question is not whether Opus 5 wins a benchmark. It is whether one workflow that failed before can now produce accepted work at a cost the business can repeat.
Start with one live workflow. Keep the finish line fixed. Measure finish rate, fixing time, review time, and cost. Keep the person who owns the work in control.
Attainment helps teams find the workflow worth diagnosing, set the human-control boundary, and build the first practical system when the gap is real.
We review fit first, then confirm scope, timing, and paid diagnostic terms before work begins.
Sources
Founder & Managing Director, Attainment
David Cyrus is the founder of Attainment. He leads the team that diagnoses the one workflow limiting an organization's growth or efficiency, then builds the strategy, AI automation, and systems to fix it, across healthcare, professional services, home services, PE-backed operators, funded organizations, and government contractors.
Connect on LinkedIn