Anthropic Won't Let You Use Their Best Model.
Prediction Markets Are Trying to Price It Anyway.

April 2026: The Most Predictable Month in AI History

April was the most intense release month the AI industry has ever produced. OpenAI shipped GPT-5.5 on the 23rd. Anthropic shipped Claude Opus 4.7 on the 16th. DeepSeek dropped V4 in open-weight form twenty-four hours after GPT-5.5. Google rolled out Gemini 3.1 Ultra. Four frontier labs, four major launches, all inside three weeks.

What's remarkable isn't the cadence. It's that prediction markets called nearly all of it.

The GPT-5.5 release-date market didn't just predict the month — it pinpointed the day. $767K traded on the question of whether OpenAI would ship on April 23 specifically. It resolved YES.

The Claude markets were even more impressive, because they correctly distinguished between two things that look identical from the outside: a release of some new Anthropic model, and a release of Claude 5 specifically. Anthropic was always going to ship something in April. They shipped Claude Opus 4.7, not Claude 5. Markets nailed the distinction:

• Claude 4.7 by April 16: resolved YES at 100% ($162K traded).
• Claude 5 by April 30: resolved NO at 0% ($1.34M traded). It peaked at 47% on April 9 and was crashing well before resolution.
• Claude 5 by May 31: hit 65% on April 9, currently sits at 11%. The market read “Anthropic ships frequently” correctly, then read “but they call it 4.7, not 5” correctly too.

This is what prediction markets do well: take a noisy stream of leaks, hiring patterns, blog posts, and rumors and price each one against the question that actually matters. The result is often surprisingly precise — accurate enough that, by April 30, traders had built a near-perfect scorecard for the entire month's release calendar.

Then Anthropic Revealed a Model You Cannot Use

On March 26, a data leak from Anthropic exposed the internal name of an unreleased frontier model: Claude Mythos. On April 7, the company confirmed it. The blog post described Mythos as “a step change” over Claude's public models and “the most capable we've built to date.” Then it announced that Mythos would not be made generally available.

Instead, Anthropic launched Project Glasswing — a preview program that gave roughly forty hand-picked organizations defensive cybersecurity access to Mythos. The list reads like a federal infrastructure roster: AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorgan Chase, Microsoft, NVIDIA, plus government partners. Their job: turn Mythos loose on the software they own and report what it found.

What it found was alarming. Within weeks of partner access, Mythos had autonomously discovered thousands of zero-day vulnerabilities across every major operating system and every major web browser. Including a flaw in OpenBSD — a project whose entire pitch is its security review process — that had gone undetected for twenty-seven years.

Anthropic's position is unambiguous. From the announcement: “We do not plan to make Claude Mythos Preview generally available.” The capability that makes it useful for defensive cybersecurity also makes it dangerous for offensive cybersecurity. The company says it intends to fold Mythos-class capabilities into a future Opus release once additional safeguards are in place — but won't commit to a date.

Markets reacted exactly the way you'd expect them to react to a glass wall. The “Mythos by April 30” market had been hovering between 25% and 39% all of late March. On April 7 — the day Anthropic confirmed Project Glasswing — it dropped from 28% to 7% in twenty-four hours. By April 10 it was at 3%. It resolved NO at 0%. $243K traded.

That collapse is the cleanest possible signal that the market understood what Glasswing was: not a precursor to a public release, but the release strategy itself. Anthropic wasn't building toward a launch. They had launched — into a forty-customer ceiling.

The Market That Refuses to Resolve

One Mythos market is still alive: “Will a Claude Mythos model be released by June 30, 2026?” It currently sits at 27%. $75K has traded.

Three things stand out about that chart. First, the baseline: traders settled in at roughly one-in-four odds within days of Glasswing. Even after the company stated explicitly that no public release was planned, the market refused to write Mythos off entirely. Anthropic does not telegraph release dates well — Claude Opus 4.7 was a surprise in name even if not in timing — and traders are pricing the residual probability that the company changes course.

Second, the spike. On April 22-23, the market jumped from 25% to 42% in two days. Those are exactly the days OpenAI shipped GPT-5.5. The mechanism is obvious: traders briefly priced in a competitive response from Anthropic, the kind of escalation pattern that has driven the whole industry for two years. If OpenAI ships, Anthropic must ship.

Then they didn't. Anthropic shipped nothing in the week after GPT-5.5. The June 30 market drifted back from 42% to 21% over the following week — the market revising its prior in the face of evidence. Anthropic is not racing. The race-or-don't-race question got answered, and the answer was “don't race.”

Third, the tail. The market closed April at 21% and is currently at 27%. Not zero. The 27% encodes everything we don't know: whether the safeguards land faster than promised, whether competitive pressure forces Anthropic's hand, whether a leak forces the timing. The point of a prediction market isn't to be certain — it's to compress every public signal into a single price. 27% is the market saying it is not certain, and refusing to pretend.

Why Anthropic Wins May Anyway (Even Without Mythos)

Here is the trader psychology question that makes this entire setup interesting. If Mythos isn't shipping, what fills the gap? Answer, from the largest active AI market on Polymarket:

That number is the actual story. Anthropic's 80% odds are not a bet on Mythos. Mythos can't be ranked on LMArena because LMArena ranks deployed models by Elo from human comparisons, and Mythos has no humans comparing it. The 80% is a bet on Claude Opus 4.7 and Sonnet 4.6 — the public Anthropic models — holding the top of the leaderboard against everything OpenAI, Google, xAI, Meta, and DeepSeek have already shipped.

Look at the structure. A trader who believes Anthropic dominates can buy YES at 80¢ and bank the spread if Claude Opus stays #1. A trader who believes Anthropic is overrated has to fade them at 80¢, which prices in a 20% probability that something replaces Claude on the leaderboard before May 31. That something has to come from somewhere: GPT-5.5 already shipped and didn't take #1; Gemini 3.5 hasn't shipped and the May 31 market gives it 23% odds; Grok isn't close. The 20% residual is essentially “Google announces something stunning at I/O on May 19-20.”

The cleanest play, if you actually believe what the market believes: buy YES on Anthropic in May, buy NO on Mythos by June 30. Both bets are pure exposure to the Opus 4.7 thesis. You win if Anthropic is dominant and they keep Mythos in the box. The fact that those two bets aren't obviously contradictory is the whole story.

And that's why the liquidity is where it is. $223K on Anthropic winning May. $75K on Mythos shipping by June 30. Traders are confident on the dominance question and uncertain on the release question — exactly the picture you'd expect when one company has decided to compete on benchmarks while withholding their best score.

What This Means for the Race

Zoom out, and Mythos is the most extreme example of a pattern that's been growing all year. Capability is decoupling from deployment. Labs are building things they don't ship for reasons that range from safety (Anthropic on Mythos) to revenue protection (OpenAI's slow rollout of GPT-5 reasoning) to regulatory caution (Google's Gemini Ultra release cadence). The leaderboards capture none of it.

Prediction markets are filling the gap. “GPT-6 before GTA VI” sat at 58.5% in March and is now at 66% — the window for OpenAI to ship is closing fast. Gemini 3.5 by June 30 is at 43%, with Google I/O on May 19-20 as the obvious catalyst.

The deeper bets are even more interesting. Polymarket gives Anthropic a 68% chance of IPO'ing before 2027, against 30% for OpenAI — despite OpenAI raising $122B at an $852B valuation in March. OpenAI's CFO publicly stated they aren't ready for a public listing this year, and the market is taking her at her word. Kalshi has the “OpenAI announces AGI” market at 11% — a useful reminder that the people putting up real money are not in the AGI-imminent camp the press releases suggest.

None of these markets give you a perfect view of the future. They give you the price of one specific bet, with all the caveats that implies. But when leaderboards are blind to the most capable models and benchmarks lag deployment by months, a price compiled from real money tends to be the cleanest signal available — even, or especially, when what it says is that nobody knows.

The First Scoreboard the Leaderboards Can't Keep Up With

The history of AI benchmarks is the history of metrics getting gamed and replaced. MMLU saturated. HumanEval saturated. SWE-bench is on its way. LMArena was supposed to be the answer — humans, blind comparisons, no training-set contamination — and for two years it mostly worked. Mythos is the first model that breaks it cleanly, not by gaming the benchmark but by refusing to enter.

Prediction markets aren't a replacement for benchmarks. They can't tell you whether Mythos is actually better than Opus 4.7; nobody outside Project Glasswing can. What they can do is price the residual question — given everything we know, what should we expect? — and hold that price up to a steady stream of news. That's what they were doing all of April, successfully, on every release date in the industry. And that's what they're still doing with Mythos, even as the model recedes further behind a wall.

Anthropic has decided that the best AI model is one you cannot use. Markets are quietly insisting that they will price it anyway.