Claude Opus 4.8 vs Claude Mythos: Released vs Preview

I’m Dora. Two people asked me the same question last week: “Should I migrate our pipeline to Claude Mythos, or wait for Opus 4.8?” Both had the framing wrong. Mythos is not a model you pick up next to Opus on the API console. Opus 4.8 is now the default-available flagship per Anthropic’s Opus 4.8 launch post; the preview sits behind an invite list. If you’re a builder evaluating either for production, this piece is about why the distinction matters before you write a single line of routing code.

No speculation about preview performance I haven’t tested.

Claude Opus 4.8 vs Claude Mythos at a Glance

Public release vs gated research preview

Two different categories. That’s the whole point.

Opus 4.8 shipped on May 28, 2026. Generally available. Standard pricing, standard API ID claude-opus-4-8, standard checkout flow. If you had an Opus 4.7 integration on Thursday morning, you had 4.8 on Thursday afternoon — same endpoint, same docs, additive changes only.

Mythos Preview is something else. Anthropic describes it as a gated research preview distributed through Project Glasswing, a defensive cybersecurity coalition. Out since April 2026, but “out” here means available to about 50 vetted organizations — AWS, Apple, Google, Microsoft, Cloudflare, CrowdStrike, the Linux Foundation, and roughly 40 others maintaining critical infrastructure. Anthropic has stated it does not plan to make the preview generally available.

So when someone compares the two, they’re not comparing products you can choose between. They’re comparing what’s shipping to you against what’s shipping to a closed list.

Builder availability and access paths

Feature	Claude Opus 4.8	Claude Mythos Preview
Status	Generally available	Gated research preview
Access	Self-serve API, claude.ai, Claude Code	Invitation only via Project Glasswing
Platforms	Claude API, Bedrock, Vertex AI, Foundry	Same four platforms, gated
Use case framing	General-purpose flagship	Defensive cybersecurity research
Production fit	Yes	No — preview, not GA

There’s no application path in the conventional sense. A Cyber Verification Program exists for qualifying security teams, but it isn’t a waitlist that converts to general access. Please refer to Anthropic’s latest official documentation for current access criteria — this policy gets updated without much warning.

What Anthropic Confirmed About Opus 4.8

API availability, pricing status, and benchmark direction

The factual surface is clean. Anthropic shipped Claude opus 4.8 at the same rate as 4.7 — $5 per million input tokens, $25 per million output — on the Claude API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. The 1M-token context window carries over (200K on Foundry at launch). I’m not going to repeat every chart from the launch post; see the Claude API documentation for Opus 4.8. The pieces that actually changed builder behavior:

Effort dial visible in claude.ai, not just an API parameter. Levels: low → medium → high (default) → xhigh/extra → max. On coding tasks, default high spends roughly the same tokens as 4.7’s default but performs better.
Fast mode as a research preview: ~2.5x output speed at $10/$50 per million tokens.
Dynamic workflows in Claude Code (research preview), orchestrating hundreds of parallel subagents.
Mid-conversation system messages in the Messages API — small but useful for long-running agents.

On the opus 4.8 benchmarks front, Anthropic reports SWE-bench Pro at 69.2% (up from 64.3%) and similar gains on agentic coding, multidisciplinary reasoning, and scientific evaluations. I haven’t re-run these on my own workload. The honesty improvement — Anthropic claims roughly 4x less likely to let flaws pass unremarked — is the one I’m curious about, because that’s the kind of regression that costs nothing on a leaderboard but real money in production. To be verified.

One pricing note buried in the announcement: fast mode is about three times cheaper than fast mode on previous models. If you’d dismissed it as too expensive on 4.7, the math is different now.

What Anthropic Confirmed About Claude Mythos Preview

Project Glasswing, defensive security use, and restricted access

The preview came out of the closet on April 7, 2026, alongside Project Glasswing. The framing is specific: defensive cybersecurity program, not product launch. Anthropic’s Project Glasswing announcement lays out the structure — twelve founding partners, around forty additional critical-infrastructure organizations, and $100M in usage credits.

What the model does, in Anthropic’s framing: identifies vulnerabilities in large codebases, builds working exploits with less manual guidance than previous models, runs long-horizon security operations. The Mythos Preview red team report walks through specific cases, including a fully autonomous remote code execution chain against an open-source NFS implementation. The capability profile is the reason for gating, not a marketing line.

Why Mythos is not a normal public model launch

Most online chatter treats the preview like Opus’s bigger sibling. It isn’t.

The capability tier is higher — Anthropic’s own alignment numbers put Opus 4.8 as “similar to Mythos Preview” on misalignment metrics, which tells you the preview was the reference point — but distribution is the bigger story. Anthropic chose to ship it narrowly to vetted partners rather than broadly at premium pricing. A deliberate choice about a frontier capability with both offensive and defensive uses. Different category from “we’re charging more for the smart one.”

Anthropic has said Mythos-class models will reach broader customers “in the coming weeks” with additional safeguards. Please refer to Anthropic’s latest official documentation for the actual rollout window — “coming weeks” is the kind of phrase that moves.

Why Builders Should Not Confuse Opus 4.8 and Mythos

Production readiness vs preview containment

This is where I see teams trip.

If you’re scoping a roadmap, the answer is almost always Opus 4.8. Standard SLAs, pricing, support. The model you build a product on.

A research preview isn’t that. Previews can change, retire, shift pricing, or have access revoked. They don’t carry the same production guarantees. Even if your org got into Glasswing tomorrow, the right way to think about it is as a tool for a specific defensive workload — not as a drop-in for a customer-facing product.

I paused here when one of those two people kept pushing back. “But if it’s better at reasoning, why wouldn’t I use it for our agentic workload?” The honest answer: “better at reasoning” and “appropriate for your product” are different sentences. The first is capability. The second is fit. They overlap less than the marketing language suggests.

API planning, model routing, and vendor risk

If you’re designing for the next 6–12 months, here’s what I’d put in the planning doc:

Build on Opus 4.8 today. Migration cost from 4.7 is near zero — same model ID family, same pricing, same context window, additive API surface.
Plan for a Mythos-class GA model. Anthropic has telegraphed it’s coming. Design your abstraction so swapping in a higher-tier flagship is a config change, not a refactor.
Don’t bet a roadmap on preview access. Even if you qualify for Glasswing, don’t make a preview-only feature your differentiator. The floor can move.
Watch alignment metrics, not just capability ones. The reported alignment improvement on Opus 4.8 — closer to Mythos than to 4.7 — matters more in production than another point on a coding benchmark.

For teams already running multi-model routing, this is a non-event. You add claude-opus-4-8 next to whatever else you route to, A/B it on real workloads, decide based on your evals. Aggregation platforms like WaveSpeedAI sit in this lane — one endpoint, switch the model string, compare. Boring, correct. Not “wait for Mythos.”

What to Watch Next

A few things I’m tracking. Not predictions — just what’s worth checking in on:

The Mythos-class GA rollout window. “Coming weeks.” If that slips, the reason matters more than the date.
The honesty / misalignment gap. Anthropic published a chart showing Opus 4.8 and Mythos effectively tied at ~1.9 on internal misalignment. If that holds in independent evals, the value of preview access drops.
Dynamic workflows in real codebases. A research preview that either lands hard or disappears. I’ll know after running it on a real migration.

This is where my data ends. Continuing next week.

FAQ

What is Claude Mythos Preview?

A frontier research model from Anthropic, distributed under a gated research preview as part of Project Glasswing — a defensive cybersecurity coalition. Access is invitation-only, prioritized for organizations working on critical software security. Not part of the public Haiku/Sonnet/Opus tier. Please refer to Anthropic’s latest official documentation for current scope and eligibility.

How is the preview different from Claude Opus 4.8?

Two categories. Anthropic Opus 4.8 is the generally available flagship — anyone with a Claude API key can use it. Mythos is a gated research model with a defensive-security framing, distributed to vetted partners only. Anthropic positions it above Opus 4.8 on its internal capability ladder, but the more important difference is distribution: one ships to everyone, the other narrowly.

Can developers access Claude Mythos through the API?

Not through self-serve channels. No signup, no waitlist that converts to standard access, no pricing page entry for general customers. The model identifier exists on Claude API, Bedrock, Vertex AI, and Foundry, but the gate is at the access layer. Qualifying security teams can apply through Anthropic’s Cyber Verification Program — possible intake path, not a guaranteed route. Please refer to Anthropic’s latest official documentation for current criteria.

Why is the preview gated?

Anthropic’s stated reason: the model’s cybersecurity capabilities — including autonomous exploit development — create a different risk profile than a typical frontier release. Gating allows Anthropic and partners to develop safeguards before broader deployment. Not a “too dangerous to release” story; an access-control decision about a capability with both offensive and defensive uses. Please refer to Anthropic’s latest disclosures for the current safety posture.

Is “opus 4.8 max” a separate model?

No. Opus 4.8 max isn’t a different model — it’s the highest setting on the effort dial. The full ladder: low → medium → high (default) → xhigh/extra → max. Higher settings let the model think longer for better quality. If you searched expecting a separate tier, you were probably looking for max effort, not a separate SKU.

Should builders plan around Opus 4.8 or wait for Mythos?

Build on Opus 4.8 today. Anthropic has said Mythos-class capabilities will reach broader customers “in the coming weeks,” but a roadmap is not a release. Design your stack so a higher-tier flagship is a config swap when it lands. Don’t bottleneck a product on preview access.

Conclusion

Two products. Two categories. Don’t confuse them in your migration plan.

Anthropic Opus 4.8 is the model you ship on this quarter — generally available, same pricing as 4.7, additive API changes, real improvements in coding and honesty. Claude Mythos is a gated research preview for defensive cybersecurity, with a path to broader availability that Anthropic has signaled but not committed to a date. Build for the released tier, design for the preview tier to land later, and you’re fine. Bet a roadmap on preview access and you’re not.

That’s it. Run it on your own workload. That’ll tell you more than I can.

Previous posts: