Claude Mythos, GPT-5.6, and DeepSeek's Pricing

Every few months, the AI release calendar turns into a pile-up—every lab suddenly shipping within the same 30-day window like they all got the same memo. We might be walking into one of those moments right now, and the signal-to-noise ratio is genuinely difficult to parse. So let me try.

The short version: Anthropic has Claude Mythos 1, Sonnet 4.8, and Opus 4.8 all in various stages of pre-release visibility. OpenAI has something internally called GPT-5.6 (model designations unconfirmed against OpenAI's public release record—treat this as a leak, not an announcement). DeepSeek just made a pricing move that should make every developer running agentic workflows stop and do the math. And June is shaping up to be either a genuinely historic month or a very overhyped one. Possibly both.

Mythos is real. Whether it's yours anytime soon is a different question.

The clearest signal here is that Claude Mythos 1 has become hard to ignore. Backend strings referencing it have appeared inside Claude Code. The model slug "Claude Mythos 1 Preview" has surfaced. It briefly became visible inside the Claude app before disappearing. It's showing up in contexts tied to Google Cloud and AWS's vulnerability discovery program—which is infrastructure-level testing, not vibe testing.

WorldofAI's coverage, citing Testing Catalog (a model-tracking publication that monitors API strings and deployment signals across major AI platforms), notes that Anthropic seems more open to a public Mythos release than their earlier communications suggested. If you've been following the Mythos leak history, you know the earlier framing was squarely "enterprise and security tooling only." That framing appears to be softening.

What we don't know: whether the public gets the full Mythos or some distilled variant. WorldofAI's read is that a distilled version arrives by October at the latest. That timeline feels plausible given where the infrastructure testing seems to be. But "plausible" and "confirmed" are doing very different jobs in that sentence.

Sonnet 4.8 and Opus 4.8: the version-skip situation

Sonnet 4.8 and Opus 4.8 model slugs were spotted on Google Vertex AI—the same kind of pre-release registry appearance that preceded Opus 4.7. The Vertex sighting alone is a meaningful signal; labs don't generally push model identifiers into cloud partner infrastructure until deployment is close.

What makes Sonnet 4.8 interesting is the backstory. According to WorldofAI's reporting, Anthropic appears to have skipped Sonnet 4.7 entirely. The evidence: references to Sonnet 4.8 showed up inside an accidentally-shipped 500K internal debugging source map that leaked through a Claude Code NPM package update on March 31st. Testing Catalog was among the first to flag the significance of what was buried in that source map—specifically, references embedded in unreleased keyword filtering systems.

The rumored spec list for Sonnet 4.8 includes visual understanding accuracy above 98% for UI mockups and messy architecture diagrams, cleaner one-pass code completions, more literal instruction following, and a new "X-high" reasoning effort tier designed to give you stronger reasoning without the generation-time penalty that current high-reasoning modes carry. I want to be clear: these figures are circulating in leak analysis, not from Anthropic directly. The 98% visual accuracy number, in particular, has no named researcher or timestamped post behind it that I can independently verify—so file it under "aspirational rumor" until Anthropic publishes evals.

The one concrete trade-off that's showing up consistently: the updated tokenizer reportedly consumes around 30% more tokens on identical prompts compared to older Sonnet models. If true, that's not a minor footnote—it means your costs go up even if the per-token price stays flat. WorldofAI didn't sugarcoat this concern: "I'm scared cuz it's an Anthropic model, and these guys have some sort of problem with rate limits, and due to the shortage of compute that they have right now, I'm not too sure and confident of any model drop from Anthropic right now." That's a fair read. Anthropic's compute constraints have been a real friction point, and a June launch of multiple models simultaneously would stress that further.

GPT-5.6: the codenames are piling up

On the OpenAI side, internal testing tags—"Iris Alpha," "Ember Alpha," "Beacon Alpha"—have been spotted in developer environments, potentially indicating multiple GPT-5.6 variants under evaluation. WorldofAI also mentions canary rollout references appearing, which matches the same pre-launch fingerprints seen before previous OpenAI releases.

Important caveat: GPT-5.6 as a model designation is entirely unconfirmed against OpenAI's public release record. I'm not linking to any prior coverage treating it as an established product because doing so would create a credibility problem the caveat can't resolve. What I can tell you is that someone who has access to the arena reportedly found the model impressive, and that the leak profile—multi-step reasoning improvements, better agentic workflow performance, improved front-end generation quality—is consistent with what you'd expect from a follow-on to whatever OpenAI's current frontier actually is.

Two variants are reportedly being evaluated: a standard and a Pro tier. The competitive context here matters: if GPT-5.6, Gemini 3.5 Pro (confirmed), and Claude Opus/Sonnet 4.8 all land within weeks of each other, the benchmark wars are going to be loud and mostly unresolvable. Every lab will have a chart where they win.

DeepSeek just changed the math

This is the part that I think deserves more attention than the model-name churn. DeepSeek has officially confirmed that the 75% pricing discount on DeepSeek V4 Pro is permanent.

WorldofAI reported specific figures—approximately $0.43 per million input tokens and $0.87 per million output tokens—but I need to flag these as unverified against DeepSeek's official pricing documentation. Third-party reporting on promotional-to-permanent pricing transitions is frequently imprecise on exact per-token figures, and you should verify current numbers directly with DeepSeek before making infrastructure decisions based on them. What is clear from DeepSeek's own announcement is that the discount is permanent, not promotional.

To understand why this matters structurally: OpenAI's flagship (whatever its current public designation) sits at $5 per million input tokens and $30 per million output tokens as of late May 2025—pricing current at time of writing, subject to change. Anthropic's Opus 4.7 runs similarly, at $5 input and $25 output. The gap between those figures and DeepSeek's confirmed permanent discount is not incremental. It's the kind of gap that changes how developers architect systems.

As WorldofAI put it: "This is quickly turning DeepSeek version 4 Pro into the new king of cost per token economics for developers building large-scale agentic systems and AI apps." For anyone running agentic workflows where token costs compound across thousands of multi-step calls, the arithmetic here is genuinely disruptive. Whether DeepSeek V4 Pro matches frontier-lab quality on your specific workload is a question you have to answer with your own evals—but the cost case for testing it is now obvious.

The pressure this puts on OpenAI and Anthropic isn't just competitive optics. It's structural: if a model in the same general performance tier costs 90%+ less, the labs charging premium prices need to either demonstrate a quality gap that justifies the delta or start moving their own prices. Neither is a comfortable position. OpenAI has been navigating this tension since DeepSeek R1 arrived—GPT-5.4's pricing story showed how uncomfortable that conversation gets.

What I'm actually watching

The Mythos situation is genuinely novel in one way: Anthropic's public posture toward releasing it has shifted. Earlier signals suggested Mythos was enterprise-and-security-only, full stop. The newer signals suggest that strong safeguards are being built specifically to enable broader deployment. That's not nothing—it means the constraint isn't capability, it's safety infrastructure. How fast that infrastructure gets built is the actual variable determining when any of us get to use this model.

The pricing war is already decided in one direction: costs are going down, and DeepSeek is the primary reason why. The open question is whether frontier quality actually tracks with frontier pricing, or whether the gap is real and measurable for tasks that matter.

June is going to answer some of this. Or it'll produce three simultaneous launch events, three sets of cherry-picked benchmarks, and a discourse pile-up that makes everything harder to read. Either way, bring snacks.

— Yuki Okonkwo, AI & Machine Learning Correspondent