Updated May 30, 2026
The latest Claude model in May 2026 is Claude Opus 4.8
That is the lookup answer. The part worth your time is the next one: how an app that does not bundle its own model list hands you a brand-new model the same day it ships, and what Opus 4.8's new 1M-context default does when it meets the limits of your actual account.
Direct answer (verified May 30, 2026)
Claude Opus 4.8
- Released
- May 28, 2026, available globally same day
- API model id
claude-opus-4-8- Replaces
- Claude Opus 4.7
- Pricing
- Unchanged from Opus 4.7 (regular use)
- Where it runs
- Claude API, Amazon Bedrock, Google Cloud Vertex AI, GitHub Copilot
Checked against the Anthropic model overview at platform.claude.com/docs. If you are reading this after Anthropic ships its next flagship, treat the date stamp above as the as-of line.
What actually changed from 4.7 to 4.8
Most coverage of a model bump reprints the same benchmark grid, so here it is once, then I will move on to the part nobody else writes about. Anthropic's own numbers for the 4.7 to 4.8 jump:
| Feature | Opus 4.7 | Opus 4.8 |
|---|---|---|
| Agentic coding (SWE-Bench-style) | 64.3% | 69.2% |
| Reasoning with tools (multidisciplinary) | 54.7% | 57.9% |
| Agentic computer use | 82.8% | 83.4% |
| Knowledge-work score | 1753 | 1890 |
| Default context window | Standard | 1M (entitlement-gated) |
| Regular-use pricing | Baseline | Same as 4.7 |
Numbers are Anthropic's published evaluations for the 4.7 to 4.8 comparison. Benchmark scores are directional; your own task mix matters more than any single grid.
The headline behavior change is honesty under uncertainty: Anthropic says 4.8 is roughly four times less likely to let a flaw in code it wrote pass without comment, and more willing to say it is not sure rather than invent a confident answer. The shipped features that matter for day-to-day agent work are dynamic workflows (one session can plan and fan out many parallel subagents), an effort dial, a faster and cheaper fast mode, and a Messages API that now accepts system entries inside the messages array so you can change the model's instructions mid-task.
The part the spec sheets skip
How a Claude Code wrapper hands you 4.8 the same day, with no app update
When Anthropic shipped Opus 4.8 on May 28, the people who felt it instantly were the ones running the model through something that does not pin a model list into its own binary. fazm is one of those: it wraps Claude Code over the Agent Client Protocol (ACP) and treats the model list as data that arrives from the running agent, not a constant baked into a release.
The mechanism is small and checkable. After a session opens, the ACP SDK reports the models the connected Claude Code build supports. fazm subscribes to that and pushes it straight into the model picker:
Because the list is dynamic, there is no fazm release in the loop. The day your local Claude Code resolves claude-opus-4-8, it appears in the picker next to whatever you were already running. You select it like any other model. And because fazm asks you to bring your own Claude Pro or Max account, that usage hits the plan you already pay for rather than a separate per-token bill.
The 1M-context default is the interesting edge case
Opus 4.8 ships with a 1M-token context window as the default. That is a real upgrade, but a default is not an entitlement. Not every account is provisioned for the 1M variant, and a model that advertises a window your plan cannot use is a turn waiting to fail. fazm handles this explicitly instead of letting you find out mid-conversation.
The bridge probes the entitlement. If Claude Code reports that the account lacks the 1M-context variant, it fires a dedicated handler with the reason 1m_context, and fazm reacts by hiding the variant and dropping your selection back to the standard-context build:
The practical effect: you still get Opus 4.8 the moment it lands, but the picker only ever offers the variant your account can actually run. The 1M window shows up for the accounts entitled to it and quietly stays out of the way for the ones that are not. This is the kind of seam that a raw CLI leaves you to discover by hitting an error; a native wrapper that reads entitlements can absorb it before it reaches you.
Why "what is the latest model" is the wrong long-term question
Anthropic upgraded the flagship twice inside two months, and teased a larger Mythos-class model "in the coming weeks." If the cadence holds, the answer to "what is the latest Claude model" has a short shelf life. What does not change as often is the question underneath it: when the next model ships, how fast can you actually use it, and does your tool keep your work intact while you switch?
That is where the wrapper choice pays off. The same dynamic-list mechanism that surfaced 4.8 will surface Mythos with no code change on fazm's side. And the reasons people reach for fazm in the first place, persistent sessions that survive a Mac restart, one-click forking of a chat into a new window with full prior context, and no auto-compacting that silently drops earlier decisions, are model-agnostic. They hold whether you are on 4.8 today or whatever replaces it next month. The model is the engine; the harness is what you live in.
Want to see Opus 4.8 running inside a persistent Mac session?
Book a 20-minute walkthrough: model picker, forking, and how the bridge handles a brand-new model the day it ships.
Frequently asked questions
What is the latest Anthropic Claude model in May 2026?
Claude Opus 4.8. Anthropic released it on May 28, 2026 and made it available globally the same day on the Claude API, Amazon Bedrock, Google Cloud Vertex AI, and GitHub Copilot. The API model id string is claude-opus-4-8. It replaced Opus 4.7 as the flagship and arrived less than two months after the prior upgrade, which is a faster cadence than Anthropic kept through 2025.
What is the exact API model id for Claude Opus 4.8?
claude-opus-4-8. Starting with the 4.6 generation, Anthropic uses a dateless id format that pins to a single fixed snapshot, so there is no trailing date stamp. On Amazon Bedrock the same model carries the provider prefix anthropic.claude-opus-4-8.
What changed between Claude Opus 4.7 and Opus 4.8?
Anthropic reports gains in agentic coding (SWE-Bench-style score from 64.3% to 69.2%), multidisciplinary reasoning with tools (54.7% to 57.9%), agentic computer use (82.8% to 83.4%), and a knowledge-work score that moved from 1753 to 1890. It also says 4.8 is roughly four times less likely to let a flaw in its own generated code go unremarked, and is more willing to flag uncertainty instead of making unsupported claims. New surface-level features: dynamic workflows (plan and run many parallel subagents in one session), an effort control, a faster and cheaper fast mode, a 1M-context default, and a Messages API that accepts system entries mid-conversation.
Did Opus 4.8 get more expensive than 4.7?
No. Anthropic states regular-use pricing is unchanged from Opus 4.7. The fast mode is the part that got cheaper and quicker (Anthropic describes it as roughly 2.5x the speed and three times cheaper than prior models). If you bring your own Claude Pro or Max account through a wrapper, you are billed against that existing plan rather than per token.
How does fazm get Claude Opus 4.8 without shipping an app update?
fazm does not hardcode a model list into its binary. It wraps Claude Code over the Agent Client Protocol, and after each session/new the ACP SDK reports the models the connected Claude Code build supports. fazm's setModelsAvailableHandler (ChatProvider.swift line 1679) feeds that list straight into the model picker via ShortcutSettings.updateModels. So the day your local Claude Code resolves Opus 4.8, it shows up in the picker. No fazm release, no re-download.
Opus 4.8 ships a 1M-context default. Will fazm use it?
Only if your account is entitled to the 1M-context variant. The bridge probes this: if Claude Code reports the account lacks the entitlement, it fires setModelEntitlementMissingHandler with reason '1m_context' (ChatProvider.swift line 1688). fazm then sets userLacks1mEntitlement, hides the 1M variant from the picker, and drops your selection back to the standard-context build so a turn never silently fails on an entitlement you do not have.
Is Opus 4.8 the last model, or is something bigger coming?
Anthropic teased a 'Mythos-class' model alongside the 4.8 launch and said it would reach all customers 'in the coming weeks.' As of this writing (May 30, 2026) Opus 4.8 is the shipped flagship and Mythos has not been released. When it does land, the same dynamic-list mechanism means a wrapper that reads models from the ACP SDK surfaces it without a code change.
Related
More on running Claude Code as a daily driver
The May 2026 Anthropic 529 outage, and the bridge-layer fix
What a 529 overloaded actually looks like to a Claude Code client, and how to absorb it without leaking a billing prompt.
Keeping Claude Code context alive across long sessions
Why context survival matters more than raw window size, and how persistent sessions change the workflow.
Auto-compacting and the token waste it hides
What auto-compacting drops, what it costs, and the case for keeping full history live for a window's lifetime.
Comments (••)
Leave a comment to see what others are saying.Public and anonymous. No signup.