Claude Sonnet 5, GPT-5.6, and What Labs Aren't

This week handed the AI press exactly what it loves: a pile of leaks, a couple of benchmark claims, and enough version numbers to fill a changelog. Claude Sonnet 5 spotted in a partner API. A GPT-5.6 voice model already rolling out to some ChatGPT users. A Japanese lab called Sakana announcing that its new orchestration system matches frontier models. And — quietly, in the background — reports that Anthropic has already trained something more capable than Mythos and restricted it before most people knew it existed.

The capability story is real and worth covering. But I keep getting pulled back to the parts of this week that aren't getting enough air.

Claude Sonnet 5: The tokenizer detail that isn't just a cost story

The WorldofAI channel flagged that the Claude Sonnet 5 model slug appeared in Anthropic partner provider systems — a pattern that, according to the video, has preceded public launches by roughly five to seven days in past cycles. (That's the channel's read of prior releases; Anthropic hasn't confirmed any timeline.) If the pattern holds, Sonnet 5 drops this week.

The capability case is solid. Sonnet is Anthropic's workhorse — the model most developers actually build on, not the flagship they demo — so an upgrade here matters more to day-to-day workflows than another Mythos-tier release would. Expected improvements include a context window potentially reaching one to two million tokens, stronger multimodal capabilities, and better handling of UI mockups and architecture diagrams. Early outputs shown in the video, including an SVG rendering of a Nintendo Switch 2 generated without a reference image, do look genuinely impressive.

Then there's the tokenizer. Anthropic appears to be moving to a new one that consumes roughly 30% more tokens on identical prompts. The video frames this primarily as a cost issue — your API bills go up. That's true, and it's worth knowing. But a 30% token increase also means 30% more detailed logs of your inputs sitting on Anthropic's servers. If you're a developer feeding proprietary code or sensitive business logic through Claude, "more tokens per query" isn't just a billing line item. It's a larger data footprint per session. Check your data retention settings and your organization's API data usage agreement before Sonnet 5 lands.

The Mythos successor: accountability questions nobody is asking

This is the story I'd actually be reporting out if I had more time this week.

According to the WorldofAI video, Anthropic has trained a model that reportedly surpasses Mythos in capability — and has already restricted it from public access. The video cites Andrew Curran, described as a reputable source on Anthropic leaks, though it's worth being precise here: WorldofAI is relaying Curran's characterization, not quoting him directly in text. What comes through in the video is a paraphrase to the effect that this successor is meaningfully stronger than Mythos, with likely improvements in long-horizon reasoning, agentic coding, and planning on complex tasks.

Whether Anthropic calls it Mythos 5.1, Mythos 6, or something else entirely is genuinely unknown. So is whether it ever reaches public release.

What the video gets right is something the broader coverage keeps missing: restricting a model from public deployment does not slow internal development. As the video puts it, "stopping models like Fable 5 or Mythos 5 from being publicly available does not slow down labs internally. In fact, it may even speed things up slightly because instead of using compute and resources to serve public users, those resources can be redirected towards training, testing, evaluations, and the next generation of models."

That's accurate. It's also the part that should make you think harder about what "safety evaluation" actually gates.

When Anthropic (or any lab) restricts a model, two things are true simultaneously: there is a genuine safety rationale, and the process by which that determination gets made is almost entirely opaque to the public. Who decides what a restricted model can and can't do? What does the evaluation framework actually test for? Who audits the auditors? These aren't paranoid questions — they're the basic accountability questions we'd ask of any institution making high-stakes decisions about what powerful tools the public gets access to.

This isn't an argument that Anthropic is doing the wrong thing. The Mythos leaks we covered earlier this year suggest the caution is grounded. It's an argument that "the lab decided it wasn't ready" is not, by itself, a complete answer — and that the infrastructure for independent oversight of these decisions barely exists yet. If you're a developer or enterprise customer building on Anthropic's stack, understanding that there are capability tiers you have no visibility into is relevant to your risk model.

GPT-5.6's voice upgrade: the data surface question

OpenAI's upcoming GPT-5.6 Pro has generated pre-release excitement primarily around coding capabilities — the video describes leaked tests of a model generating a complete first-person navigable interior in a 700KB HTML file, with multiple rooms, floor layout, and movement system, in a single session. The GPT-5.5 paradox of impressive benchmarks that don't always translate to felt improvement for ordinary users applies here too, but the coding demos are at least measuring something real.

The part I want to flag is the voice model.

OpenAI is rolling out what the video refers to variously as "BD" and "GPT-1" — the naming inconsistency appears to reflect genuine ambiguity in the source material, where it's unclear whether these are two names for the same system, a codename versus a release name, or distinct things. What is clear is the capability: this is a bidirectional, real-time voice model with an August 2025 knowledge cutoff that processes audio incrementally rather than waiting for you to finish speaking. It can follow along mid-sentence, count as you talk, and correct you in the moment.

That's a meaningfully different data-capture surface than a transcription model that waits for a complete utterance.

When a voice system processes your speech in real-time, incrementally, it's not just transcribing — it's building a running model of your conversational state. The inference layer has access to your hesitations, your corrections, your sentence fragments. That behavioral signal is richer than a finished transcript. The questions I'd want answered before using this in any sensitive context: Does OpenAI retain incremental audio buffers, or only the final transcript? What's the data retention window for voice sessions? Is real-time voice subject to the same data usage policies as text API calls, and does your ChatGPT tier affect that?

None of this is a reason to avoid the feature. It sounds genuinely useful — more fluid than current voice mode, closer to actual conversation. But "more humanlike" cuts both ways: the more a system understands your conversational flow, the more it knows about you.

Sakana Fugu: the honest benchmark read

Sakana AI, a Japanese lab, announced Fugu and Fugu Ultra this week. These are orchestration systems rather than standalone language models — they're trained to route tasks across multiple existing models and combine their outputs, similar in concept to how OpenRouter's fusion API operates. Sakana claimed benchmark performance comparable to Anthropic's Fable 5 and Mythos 5 (note: "Fable 5" appears to be an internal or community naming convention for an Anthropic model tier; it's not a widely documented public release name).

WorldofAI tested the claim directly and didn't find it credible at current capability levels. That's the honest read, and I think it's right.

The more useful comparison came from a head-to-head coding test between Fugu Ultra and Claude Opus (referenced in the video as version 4.8, which appears to be a checkpoint or leak artifact rather than a standard Anthropic release designation — treat that version number cautiously). Both were tasked with generating a Crossy Road-style game. Opus produced a better result on quality, design, and polish. It also took 79 minutes, consumed roughly 940K tokens, got stuck twice, and cost $37.85. Fugu Ultra had real problems — inverted controls, camera issues, no sound effects — but finished in 22 minutes, used about 90K tokens, and cost $7.32.

That's not a tie. Opus won the thing you ship to users. But Fugu won the thing that determines whether a development pipeline is economically viable. For a solo developer or a startup watching API costs, the delta between $7 and $38 for a comparable first draft is a decision-shaping number. You'd fix the inverted controls.

The orchestration architecture underneath Fugu is the part I'd watch. The question of whether AI systems get deployed as single massive models or as orchestrated networks of specialized ones has real implications for how data flows between components, where inference happens, and who has visibility into which part of a task. That's not an argument against orchestration — it's an argument for understanding what you're opting into when you use it.

What to actually do this week

If you're a developer: before Sonnet 5 lands, review your Anthropic API data settings and understand how the token footprint increase affects your input logging exposure. If you're feeding anything sensitive through Claude, that review was overdue anyway.

If you use ChatGPT voice mode: check your ChatGPT privacy settings and decide deliberately whether you want voice history saved. The upgrade is real; so is the expanded data surface.

On the Mythos successor: there's nothing actionable here yet, because the model isn't public. But the pattern — capability exists, access is restricted, process is opaque — is one worth tracking. The accountability infrastructure for these decisions is still being built, and how it gets built will matter more than any individual model release.

The labs are moving fast. The oversight frameworks are moving slower. Those two facts, more than any benchmark number, are what actually shape your risk model.

Rachel "Rach" Kovacs is Buzzrag's cybersecurity and privacy correspondent.