Anthropic released three model updates in five weeks. If you have not been watching the changelog closely, it is easy to miss how much shifted underneath you in that window. I put this together to make sense of what actually changed, what each model is for, and what I would actually recommend using right now.
The short version
- Claude Opus 4.8 landed May 28, 2026. A reliability upgrade to Opus, not a new tier.
- Claude Fable 5 and Claude Mythos 5 landed June 9, 2026. The first models in a new tier above Opus, called Mythos class.
- Claude Sonnet 5 landed June 30, 2026. The newest generally available model, and now the default on Free and Pro plans.
If you only take one thing from this, take the last point. Sonnet 5 is the model most developers will be using day to day, and it is a meaningful step up from Sonnet 4.6, not just an incremental one.
Claude Sonnet 5: the one that matters for most of us
Sonnet 5 is built for agentic coding, tool use, reasoning, and everyday professional work, at a lower cost than Opus. A few specifics worth knowing before you switch over:
- 1M token context window. A token is roughly three-quarters of a word, so this means Sonnet 5 can hold around 700,000+ words of conversation or code in mind at once before it starts losing track of earlier parts. In practice, that is enough to work across an entire codebase or a long document in a single session.
- 128K max output tokens. This is the cap on how much it can generate in one single response, close to 96,000 words. It is separate from the context window above, which is about what it remembers, not what it produces.
- Adaptive thinking is on by default, manual extended thinking has been removed. Older Claude models let you manually switch on a "think longer before answering" mode for harder problems. Sonnet 5 decides for itself when to slow down and reason more, based on the task, so that the manual switch is gone.
- A new tokenizer. A tokenizer is the system that breaks your text into tokens, which is how usage and cost are measured. Sonnet 5 uses a different tokenizer than Sonnet 4.6, so the exact same sentence now counts as up to 35% more tokens than it used to. This does not affect what the model can do, but it does affect anyone tracking or budgeting API usage.
- Introductory pricing of $2 / $10 per million tokens through August 31, 2026, moving to $3 / $15 after that. This means $2 for every million tokens you send in, and $10 for every million tokens it sends back, at the promotional rate. After August 31, it rises to $3 and $15. This only applies if you are using the API directly, not to a regular Claude.ai subscription.
The tokenizer change is worth sitting with for a second, because the two points above connect. Since the same prompt now costs more tokens, and Anthropic knows that, they've priced Sonnet 5 lower for now to roughly offset it. Your bill should stay close to what it is now, but once the introductory pricing ends in August, it is worth doing the math again rather than assuming your costs will stay flat.
Model ID for the API:
claude-sonnet-5
If you are on Claude Code, it is already set as the default model as of version 2.1.197.
Where Opus 4.8 fits now
Opus 4.8 was not replaced by Sonnet 5, it sits alongside it. The headline change with 4.8 was reliability rather than raw benchmark gains. Anthropic reported it is roughly four times less likely than Opus 4.7 to let flaws in its own generated code pass without flagging them, and it posted a strong score on Online-Mind2Web for browser agent tasks.
Opus 4.8 also introduced an effort parameter, letting you choose between standard, extra, and max effort depending on how much reasoning a task actually needs. That parameter carried over into Sonnet 5 as well.
My working rule of thumb: reach for Opus 4.8 on complex architecture decisions, long agentic sessions, and heavy computer use work. Reach for Sonnet 5 for the daily coding, tool use, and knowledge work that used to default to Sonnet 4.6.
The new tier: Fable 5 and Mythos 5
This is the part of the announcement that got the least coverage relative to how significant it is. Anthropic introduced a tier above Opus called Mythos class, and the first two models in it are Claude Fable 5 and Claude Mythos 5.
Both share the same underlying capabilities. The difference is safety scope:
- Claude Fable 5 is the generally available version. It includes safety classifiers that automatically route flagged requests, mainly around cybersecurity and biology, to Claude Opus 4.8 instead.
- Claude Mythos 5 does not include those classifiers. It is limited to approved organizations through Project Glasswing, invitation only, with no self serve signup.
If you are building on the API and calling Fable 5, plan for a few behaviors that were not a concern with Opus or Sonnet models: new response handling for refusals, since the classifier can decline a request before generating output; a fallback path for retrying on another model when that happens; and the fact that you are not billed for a request refused before output generation, with retries on a different model refunding the prompt cache cost of the switch.
Availability on Fable 5 and Mythos 5 has been inconsistent since launch. I am currently on the Free plan, and Fable 5 appeared in my own model picker labeled "Included until July 7," which suggests this is a limited trial window rather than standing access. There were also reports shortly after launch of a temporary suspension tied to an export control directive. If your work depends on either model, I would treat availability as unconfirmed until you check it directly, rather than assuming day one access holds.
Quick comparison
| Model | Released | Tier | Best for |
|---|---|---|---|
| Claude Sonnet 5 | June 30, 2026 | Sonnet | Daily coding, agents, knowledge work, lower cost |
| Claude Opus 4.8 | May 28, 2026 | Opus | Complex architecture, long agent sessions, browser/computer use |
| Claude Fable 5 | June 9, 2026 | Mythos (public) | Hardest coding and long horizon agentic work, with safety routing |
| Claude Mythos 5 | June 9, 2026 | Mythos (restricted) | Same as Fable 5, approved orgs only, no classifiers |
What to actually do this week
If you are already on Sonnet 4.6, move to Sonnet 5. Pricing is close to cost neutral during the introductory window, and the agentic gains are real rather than marginal.
If you are running production workloads on Opus for cost reasons, benchmark Sonnet 5 against your actual tasks before assuming you still need Opus. Anthropic's own numbers show Sonnet 5 slightly outperforming Opus 4.8 on some knowledge work evaluations, which was not the case with earlier Sonnet generations.
If Fable 5 or Mythos 5 shows up as an option for you, treat it as something to evaluate rather than something to build a critical workflow on, at least until access stabilizes.
Anthropic's release cadence is not slowing down, and the gap between a new model shipping and it becoming the default everywhere keeps shrinking. It is worth checking the changelog more often than feels necessary right now.
I am Tobore, a full-stack developer and Google I/O 2026 Writing Challenge winner. I write about tools that actually change how developers work. More at dev.to/toboreeee.
Top comments (1)
The tokenizer/pricing interaction is the part I'd make every API team model before switching: a 1M context window sounds like pure upside, but "same prompt can count up to 35% more tokens" changes the economics once the $2/$10 intro rate becomes $3/$15 after August 31. I also like the practical boundary you drew between Sonnet 5 for daily coding/tool use and Opus 4.8 for long agentic or architecture work. Founder/engineer lesson here: model choice should be treated like infra config, with task benchmarks and cost alerts, not a one-time preference in someone's editor.