Where Are We With Claude Mythos? (And Do You Actually Need It?)

The story broke on March 26 when an internal Anthropic draft blog post ended up in a publicly accessible data store. The post was scrubbed within hours. Screenshots had already gone everywhere. Anthropic has confirmed the broad strokes since.

Here is the part that is real:

Property	Claude Mythos 5
Parameters	~10 trillion total (Mixture of Experts)
Active per call	A fraction of total, MoE architecture
Context window	4M tokens functional, via "tiered attention"
Tier	New "Capybara" tier, sits above Opus
Availability	Project Glasswing only, ~50 vetted orgs
Approved use	Defensive cybersecurity work
Public API access	None

The interesting technical bit is "tiered attention." Anthropic kept inference cost manageable by maintaining different attention resolution at different points in the context window. So a 4M token context does not actually require 4M tokens of full attention math. It is a real architectural innovation, not just a number on a marketing slide.

The interesting business bit is the Capybara tier. Anthropic added a model class above Opus, which means the existing pricing ladder (Haiku, Sonnet, Opus) is now four tiers deep. Even if Mythos never opens up to the public, the ladder itself is a signal. Anthropic is telling its enterprise customers that there is a "more powerful than Opus" tier coming, eventually.

The official line is safety. Mythos is restricted to "defensive cybersecurity work" through a controlled program called Project Glasswing. Selected orgs get access. Everyone else waits.

The unofficial reading, depending on who you ask, is one of three things:

One: it is genuinely dangerous in the wrong hands. A 10T-parameter model with 4M context that is good at code and reasoning is also good at finding zero-days, writing exploits, and chaining attacks. Anthropic's RSP (Responsible Scaling Policy) commits them to delaying release if a model crosses certain capability thresholds. This is them following their own policy. If you believe the policy is real, this is the simplest explanation.

Two: it is about competitive positioning. Letting AWS, Microsoft, and Google have Mythos but not the rest of us creates a dependency loop. The hyperscalers get a model their competitors cannot offer. Anthropic gets a deeper integration story with the cloud providers it depends on for compute. Everyone wins except devs.

Three: it is not actually ready. A 10T MoE with novel attention architecture is hard to serve at scale and even harder to serve at consumer prices. "Critical infrastructure only" is a face-saving way to limit volume while they figure out the inference economics. The cybersecurity framing is the cover.

The honest answer is probably some of all three. None of them are reasons for you to put your roadmap on hold.

Sit with this honestly for a minute.

What are you actually building? Most of what gets built with Claude in the wild looks like one of these:

• A coding assistant or agent that calls the API a few times per task
• A document processor that summarizes, classifies, or extracts from text
• A chatbot or support agent
• An internal tool that wraps a model around some company data
• An MCP server that exposes a system to Claude (or to Claude Code)

For all five of those, you do not need Mythos. You will not see the difference. Claude Sonnet 4.6 will handle the work. Opus 4.7 will handle it better when the task needs deeper reasoning. The marginal capability gain from a 10T model is real but it shows up in places most apps never reach: multi-step research over thousands of pages, novel scientific reasoning, adversarial security analysis. If your prompt fits in 32K tokens and the answer is one of a few well-defined shapes, a frontier model is overkill, and Mythos is overkill on top of overkill.

The honest test: can you list three concrete tasks your current product fails on, where the failure is "the model was not smart enough" rather than "the prompt was bad" or "the context was too short"? If you cannot, you do not need Mythos.

The good news is the rest of the lineup got dramatically better in the last six months.

For coding

Claude Opus 4.7 is the current ceiling for general-purpose code work. Better than Opus 4.6 on long refactors and multi-file edits, and the Claude Code integration is the smoothest agent loop available right now. Cost is real ($15 input / $75 output per million tokens), so use it for the hard parts, not for everything.

GPT-5.5 shipped a week after Opus 4.7 and is genuinely competitive on coding. OpenAI's pitch is "better at coding, computer use, deeper research," and the deeper research piece is the part that holds up. For long, exploratory coding tasks where the model needs to read a lot before it writes, GPT-5.5 is worth a head-to-head against Opus on your actual workload.

DeepSeek V4-Pro is the wild card. Open weights, just hit #1 on the Vibe Code Benchmark, undercuts everyone on price. If you are building something where you need to control the model end-to-end (or you are running into rate limits on the closed APIs), V4-Pro is the model to test next. Licensing and geopolitics are real considerations, but the model itself is the real deal.

For daily driver work

Claude Sonnet 4.6 is still the best daily-driver model in the lineup. It handles 90% of what people use Opus for, at a fraction of the cost. If you are building a product that calls Claude on every user request, Sonnet is almost always the right default until you have a specific reason to upgrade.

For bulk and budget

Gemini 3.1 Flash-Lite at $0.25 per million input tokens is absurd value for high-volume work. Claude Haiku 4.5 is faster and better at structured output. Pick based on whether you need speed or structure.

I went deep on the pricing landscape in The AI Price War Is Getting Ridiculous if you want the full breakdown.

For agents

If you are building an agent, the model matters less than your tool design and your context discipline. Sonnet 4.6 plus a well-designed MCP server outperforms Opus 4.7 with a sloppy one almost every time. I covered this pattern in MCP Tutorial: Build AI Agents That Actually Do Things.

To be fair to the "I am waiting for Mythos" crowd, there are real use cases where the wait makes sense. They are narrow.

You are doing offensive or defensive security research at a level the current models cannot reach. If your job is finding zero-days in large codebases or analyzing malware families across millions of samples, the capability gap between Opus and Mythos is reportedly meaningful. You are also exactly who Project Glasswing is letting in, so apply.

You are doing scientific reasoning across genuinely large corpora. Drug discovery, materials science, theorem proving across whole math libraries. The 4M token effective context plus the reasoning depth is the combination that matters here.

You are building something where "smart enough" is a binary, not a gradient. Some systems need the model to either solve the problem or not solve it. Most products do not work this way. They tolerate "mostly right" output. If yours genuinely does not, and you have proven that Opus is not enough, then the wait is justified.

For everything else, "I am waiting for Mythos" is what people say when they do not want to ship. Ship now. Swap models later. The API contracts are stable enough that the migration cost is small.

Three things to actually watch, regardless of whether Mythos ever opens up.

One: the tiering above Opus is here to stay. The Capybara tier is now part of the public language. Even if Mythos itself stays in Glasswing, the next model after it is going to slot into the same position, and the one after that. The "frontier" is fragmenting into "frontier you can buy" and "frontier you cannot." Plan your pricing assumptions accordingly.

Two: the cybersecurity-only carveout is going to spread. If the Anthropic precedent holds, every frontier lab is going to start carving out specific capabilities for specific sectors. This is regulatory adaptation in real time. It also creates a new procurement category, "approved AI partner for [industry]," that did not exist a year ago. If you sell into a regulated industry, this is going to matter for you within 18 months.

Three: the gap between "what is possible" and "what you can ship with" is widening, then narrowing, then widening again. Mythos exists. You cannot use it. In six months, the open-weight models will probably catch up to where Opus 4.7 is today. In nine, the closed labs will release the next tier. The gap is not stable. Building against the assumption that "the model I have today is the model I will have next quarter" has been wrong for two years and is going to keep being wrong.

The play is not to wait. The play is to build something that gets better when the models get better, without having to be rewritten when they do.