Over the weekend, a screenshot started making the rounds on X: a Vertex AI error log containing a model ID that doesn’t officially exist. claude-sonnet-5@20260203. The AI community did what it does best - ran with it.
Within 48 hours, the leak had spawned benchmark claims, pricing speculation, a codename (“Fennec”), and the discovery that a second unreleased model - Opus 4.6 - appears to be sitting on the same servers. Anthropic hasn’t said a word.
Here’s what the evidence actually shows, and where the speculation gets ahead of itself.
The Hard Evidence: 403 vs. 404
The most technically interesting part of this story isn’t the original screenshot. It’s what happened after.
A developer named Ben Taleb Jr. ran a monitoring script against Google Cloud’s Vertex Model Garden endpoints, testing model IDs to see which ones actually exist on the server. The results:
claude-sonnet-99(a fake ID): 404 Not Found - the resource doesn’t existclaude-sonnet-5(the leaked ID): 403 Forbidden - the resource exists, but access is denied
In REST API terms, a 404 means “nothing here.” A 403 means “something’s here, but you can’t have it yet.” The same test returned 403 for claude-opus-4.6, suggesting Anthropic has at least two unreleased models staged on Google’s infrastructure.
This follows Anthropic’s established versioning pattern - Opus 4.5 is claude-opus-4-5@20251101 - and the date string 20260203 in the leaked model ID points to February 3, 2026 as either a build date or a checkpoint.
The Codename: Fennec
Industry sources and early testers have attached the codename “Fennec” to the model. Anthropic has a history of animal-themed internal names for its models, so this tracks.
TestingCatalog claims to have gotten hands-on with a non-thinking Sonnet 5 variant. Their early impressions: math performance competitive with frontier models, coding output stronger than Opus 4.5 in some workflows, and particularly impressive results on structured visual generation tasks like ASCII art and UI rendering code. The tested variant had a 128k context window, though leaked specs suggest the release version could go up to 1 million tokens.
The Speculation: Pricing and Performance
This is where things get slippery. Claims circulating online include:
- 82.1% on SWE-Bench - This would beat Opus 4.5’s 80.9% launch score while sitting at a lower pricing tier. No independent verification.
- 50% cheaper than Opus 4.5 - Would put pricing around $7.50 per million output tokens versus Opus 4.5’s $75. Extraordinary claim, no source beyond community speculation.
- “Dev Team” mode - Allegedly enables parallel sub-agent generation. Worth noting: Claude Code already supports sub-agents, so this could be a marketing rebrand of existing capability rather than something new.
- 1 million token context window - Plausible given Anthropic’s infrastructure investments, but contradicted by the 128k window seen in the tested variant.
As one analysis put it: the pricing claim has “no source for this. Would be incredible if true.”
The Timing: Super Bowl Week
The leak’s timing isn’t accidental - or at least, the attention it’s getting isn’t. Super Bowl LX is this Sunday, February 8. Anthropic already purchased Super Bowl ad spots for Claude, making it the second AI company to do so after OpenAI’s ChatGPT ad last year.
Dropping a major model release during the same week you’re running your first Super Bowl commercial would be a logical marketing play. And with the Cowork product launch still reverberating through markets after triggering a $285 billion selloff, Anthropic has the attention.
It’s also only 10 weeks since Opus 4.5 shipped in November 2025, which would be an unusually short gap between major model releases.
What Anthropic Has Said
Nothing. As of February 5, there has been no official announcement, blog post, press release, or documentation from Anthropic regarding Claude Sonnet 5, its specifications, or its release date. The company has not commented on the Vertex AI leak.
What This Means
The 403/404 evidence is the strongest signal here. Anthropic clearly has at least one, possibly two unreleased models staged on Google Cloud infrastructure. That’s not unusual - companies stage models well before public release - but it does confirm that something is coming.
The rest is noise until Anthropic speaks. The benchmark claims lack independent verification. The pricing speculation has no sourcing. The “Fennec” codename is plausible but unconfirmed. And the tested variant may represent an early build that differs significantly from whatever ships.
For developers and businesses currently building on Claude: nothing changes today. Opus 4.5 and Sonnet 4.5 remain the production models. If Sonnet 5 does deliver Opus-tier performance at a substantially lower price, it would be a significant shift in the cost-performance equation for AI applications - but that’s a big “if” built on unverified screenshots.
What to Watch For
- Anthropic’s official channels - The company tends to announce models via blog posts and API documentation simultaneously
- The Super Bowl commercial - If it references a new model, that’s your launch signal
- Vertex AI and AWS Bedrock - New model availability on cloud platforms often precedes or coincides with official announcements
- Pricing confirmation - The 50% cost reduction claim is the most consequential. If true, it would force a repricing conversation across the entire AI API market