AI-First
Anthropic released Claude Opus 4.8 on May 28, 2026, just 41 days after Opus 4.7. An unusual pace, explained by a simple reason: Opus 4.7 had real-world problems that professional users could no longer ignore. I use Claude Code every day with my SMB clients, and I can tell you this update is anything but cosmetic.
What makes Opus 4.8 interesting for an SMB isn't some abstract benchmark. It's a model that owns up to its mistakes, an effort slider that lets you control your bill, and dynamic workflows that launch dozens of agents in parallel. In concrete terms, the cost of fast mode has been cut by a factor of three, and the model lets four times fewer silent bugs slip through in its own code.
- 🎯 Improved honesty: Opus 4.8 flags its doubts instead of making things up.
- ⚡ Fast mode at 3x lower cost: $10/MTok input, $50 output, 2.5x faster.
- 📊 5-level effort control: you choose how much Claude thinks (and how much it costs).
- 🏗️ Dynamic workflows: up to 100 sub-agents running in parallel for large-scale projects.
Why Opus 4.7 was a problem in an SMB context
What real-world flaws were affecting day-to-day work?
Opus 4.7 didn't deliver on its promises for teams relying on Claude as a production tool. According to InformatiqueNews, the model "would argue to the point of hallucination, resist corrections, and in some areas produce lower-quality code than Opus 4.6." For a solo developer in a 15-person company, an assistant that refuses to admit its mistakes is worse than no assistant at all.
The underlying issue is trust. An overconfident model creates invisible technical debt. I've seen this scenario play out with several clients: the code generated by Opus 4.7 passed unit tests, but logical bugs stayed buried. Review time increased instead of decreasing, which cancelled out the expected productivity gains.
The early tester community documented unstable behaviors: inconsistent answers from one session to the next, rising costs with no proportional improvement, and a tendency to state errors with full confidence. For SMBs paying a Claude Pro or Max subscription at several hundred euros per month, this kind of regression translates directly into lost hours.
What Opus 4.8 fixes (and what the benchmarks don't tell you)
Why does model honesty matter more than scores?
Anthropic took an unusual communication angle for this launch: instead of leading with benchmarks, the company emphasized the model's honesty. According to the official Anthropic page, Opus 4.8 "flags its uncertainties more often and makes fewer unsupported claims." In practice, the model is roughly four times less likely to let a bug slip through in its own code without warning you.
For an SMB, this is the most important change. An AI assistant that says "heads up, I'm not sure about this part" saves you 30 minutes of debugging. An assistant that confidently states everything works costs you three hours.
The benchmarks remain solid nonetheless. Opus 4.8 reaches 69.2% on Agentic Coding (up from 64% for the previous version and 58.6% for GPT-5.5), and 83.4% on OSWorld, the test that measures the ability to drive a browser end to end. That said, GPT-5.5 keeps the lead in Terminal Coding with 78.2% versus 74.6% for Opus 4.8. No single model is the best at everything, and that's precisely what Anthropic now acknowledges.
| Benchmark | Opus 4.7 | Opus 4.8 | GPT-5.5 | Trend |
|---|---|---|---|---|
| Agentic Coding | 64.0% | 69.2% | 58.6% | ↑ +5.2 pts |
| Terminal Coding | 71.2% | 74.6% | 78.2% | ↑ +3.4 pts |
| OSWorld (Computer Use) | 78.0% | 83.4% | 78.7% | ↑ +5.4 pts |
| Knowledge Work | 1,710 | 1,890 | 1,769 | ↑ +10.5% |
| Financial Analysis | 49.1% | 53.9% | 51.8% | ↑ +4.8 pts |
SOURCE: Anthropic, official Opus 4.8 page · Updated 05/2026
How should an SMB interpret these numbers?
These scores measure complex agentic tasks. In plain terms: the model's ability to chain multiple actions without human oversight. For an SMB, the Agentic Coding score means Claude can enter a real codebase, identify a bug, and fix it on its own in 69% of tested cases. The OSWorld score means it can drive Excel, fill out a web form, send an email, and chain these tasks the way a human would.
The most telling improvement is in Knowledge Work (1,890 points versus 1,710). This benchmark measures the ability to read documents, cross-reference information, and produce a synthesis. That's exactly the kind of task a CFO or COO delegates to an AI assistant: analyzing a contract, summarizing a quarterly report, comparing vendor proposals.
Effort control: the lever SMBs have been waiting for
How does the effort slider work?
Before Opus 4.8, the only way to control Claude's "depth of thinking" was through the API and the budget_tokens parameter in extended thinking, according to the Décodeur IA guide. Out of reach for a non-developer. With Effort Control, a five-position slider appears directly in claude.ai, Cowork, and Claude Code: Low, Medium, High (default), Extra, and Max.
The principle is simple. The higher the effort, the longer Claude thinks, the more tool calls it chains, and the more tokens it burns. The lower the effort, the faster the answer arrives and the less it costs. For an SMB managing a monthly AI budget, this slider is a real optimization lever.
Which level should you pick for each task?
My field experience with SMB clients gives me a simple framework. Low works for volume tasks: classifying 200 support tickets, triaging an inbox, rewriting product descriptions. The quality is "acceptable," not "excellent." Medium covers 80% of daily use cases: drafting sales emails, meeting summaries, document briefs. High (the default) is the right call for anything involving code, financial analysis, or client-facing content.
Above High, the Extra and Max levels are for specific situations: codebase migrations, security audits, complex legal analysis. The cost per request can double or triple. For most SMBs, I recommend staying between Low and High, and only stepping up to Extra for high-stakes tasks.
The real win is no longer paying top dollar for simple tasks. Before Opus 4.8, every request consumed the same thinking budget, whether you were asking for a "yes/no" or a 50-page analysis. With Effort Control, a 30-person SMB can easily cut its Claude bill by 20 to 30% by using Low for triage and drafting.
Dynamic workflows and Goal mode: when AI works on its own
What do dynamic workflows enable?
This is the feature that generated the most buzz in the technical community. Dynamic Workflows allow Claude Code to launch 10 to 100 sub-agents in parallel to tackle a large-scale problem. In practice, you set a high-level objective ("migrate this database," "refactor this 100,000-line module") and Claude orchestrates a team of agents that work simultaneously on different parts of the problem.
Before Opus 4.8, this kind of orchestration existed through manual setups (harnesses, custom scripts, complex configurations). The difference is that orchestration is now built in. You no longer need to understand the underlying architecture to benefit from it. According to Frandroid, dynamic workflows are limited to Claude Code for now, not yet available in the general web interface.
Should you get excited about 100 parallel agents?
No. And this is a point I want to emphasize because it echoes my conviction about AI integration in SMBs: a powerful tool used poorly creates more problems than it solves. Denny Weber, a German content creator who tested Opus 4.8 for a week, puts it bluntly: "100 agents aren't automatically better than one that actually thinks." For a large migration, yes, dynamic workflows are relevant. For an SMB's day-to-day work, Goal mode (a single agent working autonomously until the task is done) remains more reliable.
Goal mode is the other new addition. You define an objective and a budget, and Claude handles it on its own. "Get all Auth module tests passing, fix the lint, and merge cleanly." You close the laptop, come back, and it's done. This is the scenario my SMB clients have been asking about for a year: an AI that executes specific tasks while the team focuses on the business.
"The value of AI for an SMB isn't 100 agents running in parallel. It's one agent that correctly handles a specific task, start to finish, without supervision."
Vincent, June 2026
What it costs (and why you shouldn't wait for Mythos)
Is fast mode cost-effective for an SMB?
Opus 4.8's fast mode runs 2.5 times faster and costs $10/MTok input, $50 output, according to the official Anthropic documentation. Three times cheaper than the fast mode on previous versions. The standard mode price remains identical to Opus 4.7 ($5/MTok input, $25 output), which is notable in a context where AI infrastructure costs are surging.
Opus 4.8 remains the most expensive model on the market per token. GPT-5.5 and Gemini 3.1 Pro cost less. But if Opus 4.8 solves a problem in 3 tool calls where GPT-5.5 needs 7, the total cost flips. Cursor testers have measured it: "tool calling is significantly more efficient, with fewer steps for the same intelligence."
My advice for SMBs weighing Claude against GPT aligns with what I explain in my article on real-world Claude use cases for businesses: don't compare per-token prices, compare cost per completed task. With Effort Control, Opus 4.8 lets you bring costs down on simple tasks without switching providers. That's an argument GPT-5.5 can't match.
Should you wait for Mythos before investing?
Anthropic slipped in a line at the end of its announcement: Mythos will arrive "in the coming weeks" for all users. The model is already available in limited access through Project Glasswing. My answer is unequivocal: adopt Opus 4.8 now. Mythos will be more expensive, slower, built for extreme use cases. SMBs that wait for the "best possible model" before making a move never make a move. The value isn't in the model itself, it's in the integration with your business processes.
Frequently Asked Questions
Is Claude Opus 4.8 available on all subscription plans?
Yes. Opus 4.8 is accessible on Claude Pro, Team, and Max via claude.ai, as well as through the Claude API, Amazon Bedrock, and Vertex AI. The effort slider is available on all plans. Dynamic workflows are currently limited to Claude Code.
Does Effort Control really reduce costs?
Yes, provided you use it wisely. By running triage, classification, and drafting tasks in Low mode, an SMB can reduce its token consumption by 20 to 30% with no visible impact on quality. The savings are even more significant for teams processing a high volume of repetitive requests through the API.
Does Claude Opus 4.8 replace GPT-5.5 for an SMB?
Not across the board. GPT-5.5 remains stronger in Terminal Coding (78.2% versus 74.6%) and can cost less per token. Opus 4.8 leads in Agentic Coding, Computer Use, and Knowledge Work. The right choice depends on your primary use cases. For agentic coding and document analysis, Opus 4.8 is ahead. For pure terminal execution, GPT-5.5 keeps the edge.
Are dynamic workflows useful for a 20-person SMB?
In most cases, no. Dynamic workflows shine on large-scale migrations (100,000+ lines of code) and massive refactors. For an SMB, Goal mode (a single autonomous agent that works until the task is resolved) covers nearly all needs. Save dynamic workflows for exceptional technical projects.
Should you wait for Claude Mythos before investing in Claude?
No. Mythos will likely be more expensive and designed for extreme research or analysis tasks. Opus 4.8 offers the best value for SMB use cases as of June 2026. Integrating AI into your business processes matters more than raw model power.
Vidéos YouTube
- I tested Opus 4.8: here is what you need to know · Benjamín Cordero
- DIESE 7 Dinge ändern jetzt ALLES mit Opus 4.8! · Everlast AI
- Giải Mã Benchmark CLAUDE Opus 4.8 · TOMMY academy
- Eine Woche Opus 4.8: Comeback oder teurer Reinfall? · Denny Weber