Forecasting Claude Leaks
Will ClaudeCode get an upgrade today in the form of Sonnet 5?
Whispers and Rumors
Twitter and Reddit are awash with rumors about Anthropic’s latest offerings. Sonnet 5, supposedly codenamed “Fennec,” was due to be released yesterday, or perhaps today, or maybe later this week or month or year. And we might also be getting Opus 4.6 at the same time?
I know that it’s all most of you can do to stop yourselves from selling your hair and teeth to gain access to the latest models just one hour earlier, so in uncertain times like these, we turn to the wisdom of the crowds.
Traders were pretty confident through most of Monday and Tuesday that Sonnet 5 was due to be released around 10am PT on Tuesday. When it wasn’t, the market on Sonnet’s release this week took a nosedive, but recovered on robust reports that perhaps today (Wednesday) would be the day the model is released instead. Now, as the hours creep on with no hint of a new release from Anthropic, the market is slowly ticking down, and likely to crater if morning on the West Coast passes with no release.
Traders have staked some pretty massive sums on the outcome of this market, with Manifold’s #2 trader, Bayesian, pairing up with a less profitable but more bread-knowledgeable trader, Mary Smith on the YES side, against Manifold sharps Walmart and It’sMe. Whether or not Anthropic tweets out an update in the next few hours might make or break some Manifold fortunes.
Traders still think the odds are slightly better than a coin flip that Sonnet 5 will be announced today. Traders seem to have been caught off guard earlier this week, both by the early release date, but also that Sonnet 5 would be released without an intermediate 4.6 release. A Sonnet 4.6 release is still remotely possible if the persistent rumors around the release end up being some form of mass hysteria.
It also seems likely that Claude Opus 4.6 will be released this month, not far behind Sonnet 5, with ~70% odds of a release in February, and the largest YES holder being one of Manifold’s top traders, holding court against a brave user who has put about half their net worth on NO.
New Capabilities
Why does any of this matter? Well, ClaudeCode currently represents the cutting edge of AI capabilities, according to most of the AI world’s intelligentsia (and even representatives of Google and OpenAI who acknowledge they’re playing catch-up). Releases that push that frontier even further will be met with tremors on Wall Street and gratitude from the millions of startups, SWEs, and developers who are now relying on ClaudeCode for rapid productivity improvements.
You can also, of course, bet on some of the capabilities of the imminent Sonnet 5 in a props market. Traders think it might come with “Agent Swarm” capabilities, which Twitter rumors describe as releasing multiple teams of synchronized agents simultaneously, in a “swarm,” upon your request. Traders also think it likely that the context window for Sonnet 5 will improve, and that it will outperform it’s larger and three month-older sibling, Opus 4.5 on WebDev arena and in the eyes of Manifold poll respondents.
Traders also think that Anthropic is due to release an image model soon, perhaps in conjunction with these upcoming releases.
However, traders are believing rumors that the image model will likely have not been developed in-house, but outsourced, perhaps to Google or another AI lab. This is in line with Anthropic historically not prioritizing image/video capabilities but focusing on writing and code-generation. However, in a world where many are using ClaudeCode to coordinate subagents which interface with other AI models to generate images, it’s probably in the interest of Anthropic to streamline that with a more direct partnership or the development of their own image-gen model.
Forecasters expect Sonnet 5 to have a slightly higher METR task length time horizon than even Opus 4.5, currently the state-of-the-art at around 6 hours. METR has continued to see faster and faster doubling-time estimates, and a big phase change has been the success of agentic scaffolding through the continued development of ClaudeCode and other command line interfacing capabilities. Increases in time horizon are not only tied to more intelligent models (the smaller Sonnet 5 may be slightly dumber than Opus 4.5) but also tweaks to these models’ ability to fit in seamlessly with skills, hooks, memory, pipelines, and other quirks of CLI utilization.
Traders expect that when Opus 5 comes out later this year, there’s a good chance it will require METR to expand their task portfolio to include more tasks above 10 hours, which are currently lacking.
As for the more subjective metric of whether Claude power users will prefer Sonnet 5 to Opus 4.5, Manifold traders give this pretty good odds, at around 70%.
Let’s hope the power users have time to make up their minds, as it looks to be a busy February, with other models such as GLM 5, GPT-5.3, Gemini 3 Pro (non-preview), and perhaps even a Grok 4.20 likely to come out over the next 3.5 weeks.
Buy/Sell/Hold Advice
While many traders prefer anonymous leaks and rumors to serious expert advice, I asked recently crowned World #1 Forecaster Peter Wildeford for his advice to casual traders for a few of these Manifold markets (advice for reputational purposes / play-money transactions only), after he correctly predicted on Tuesday that Sonnet would not be released.
He recommended BUY on Sonnet 5 being released in February, HOLD on a release this week (when the market was at ~70%), BUY for a release today (at ~50%), SELL on a simultaneous Opus 4.6 and Sonnet 5 release (~20%), and SELL on a Claude Opus 4.6 release in February (~70%).
Take those hot Mana tips as you will, and Happy Forecasting!
-Above the Fold