The Signal — April 24, 2026

Today's stories share a thread: the gap between what AI companies say their technology is and what the people building, regulating, and using it actually experience.

OpenAI Launches GPT-5.5, Doubles API Pricing, Calls It "A New Class of Intelligence"

OpenAI released GPT-5.5 today, codenamed "Spud," because apparently even paradigm-shifting intelligence needs a folksy nickname. The model is available to paid ChatGPT and Codex users, with a Pro tier rolling out for Business and Enterprise customers. OpenAI is calling it "a new class of intelligence," which is the kind of phrase that either ages like wine or like milk, with very little middle ground.

The benchmarks are genuinely strong. GPT-5.5 hits 82.7% on Terminal-Bench 2.0 (state-of-the-art) while matching GPT-5.4's latency, meaning it's more capable without being slower. It beats Claude Opus 4.7 and Gemini 3.1 Pro across most evaluations. It also uses fewer tokens per task, a practical efficiency gain that matters more to developers than any benchmark number. The competitive positioning is clear: OpenAI wants this to be the model you compare everything else against.

The catch is the price. API costs double compared to GPT-5.4. OpenAI is betting that improved capability and token efficiency offset the sticker shock, but doubling the price of your flagship product is a bold move when Anthropic and Google are iterating on competitive models at aggressive price points. For enterprise customers locked into OpenAI's ecosystem, the upgrade path is straightforward. For everyone else shopping between frontier models, this is a value calculation that just got more complicated.

The broader question is whether "new class of intelligence" means anything concrete or whether it's marketing language for incremental-but-real improvements across the board. OpenAI has earned a pattern of overpromising on vibes and delivering on benchmarks. GPT-5.5 looks like it continues both traditions.

Sources: OpenAI, TechCrunch, CNBC, Fortune, The Decoder

Trump Administration Quietly Walks Back Anthropic Hostility

After months of escalating friction (Pentagon contract blacklisting, "woke AI" rhetoric from administration officials, and a lawsuit from Anthropic against the DoD), the Trump administration appears to be softening its stance toward Anthropic. Politico reports that lobbyists say the tone has shifted, and Trump himself told reporters that a Defense deal is "possible."

The catalyst appears to be Mythos. Anthropic's cybersecurity model has demonstrated capabilities that are hard to ignore when you're running the world's largest defense apparatus. The model has uncovered thousands of critical vulnerabilities across major operating systems and browsers, and the national security implications of not having access to that capability seem to have outweighed whatever ideological objections the administration was nursing.

This is a pattern worth watching beyond Anthropic specifically. The "woke AI" framing was always more about political positioning than technical reality, and it was always going to collide with the fact that these companies build things the government needs. The question was never whether the administration would come around, but what it would cost Anthropic in concessions and what precedent the whole episode sets for how future administrations use procurement as a political weapon against AI companies.

For now, the detente is informal. No contracts have been reinstated, no formal policy has changed. But the rhetorical shift matters because it signals to the rest of the federal bureaucracy that engaging with Anthropic is no longer a career risk. That's how procurement actually works: not through executive orders, but through the risk calculus of the people writing the RFPs.

Sources: Politico, The Hill, The Hill (newsletter), Yahoo News

DeepMind's Decoupled DiLoCo Makes Training Runs Self-Healing

Google DeepMind published a paper on Decoupled DiLoCo, a distributed training method that splits large training runs into asynchronous, fault-isolated "islands" of compute. The headline number: 88% goodput even under high hardware failure rates. For context, a typical large training run can lose days of compute to a single node failure. DiLoCo treats failures as expected events rather than catastrophes.

The technical approach is straightforward in concept, if not in execution. Instead of running one massive synchronized training job across thousands of GPUs where any failure can stall or corrupt the entire run, DiLoCo partitions the work into independent groups that synchronize periodically. If one island fails, the others keep going, and the failed island reintegrates once it recovers. DeepMind tested this with Gemma 4 models across geographically distant data centers.

The practical implications are significant for anyone trying to train frontier models. GPU clusters fail constantly at scale (it's not a question of if but how often). A method that maintains 88% efficiency despite those failures directly translates to faster training timelines and lower costs. It also opens the door to training across multiple data centers that don't need to be co-located, which matters as the industry runs into physical limits on how many GPUs you can cram into one building.

This is the kind of infrastructure paper that doesn't generate headlines but shapes what's possible twelve months from now. If DiLoCo or something like it becomes standard, the constraint on frontier training shifts from "can we keep a cluster running long enough" to "can we get enough total compute," which is a much more tractable problem.

Sources: DeepMind Blog, arXiv, MarkTechPost, The Rift

Anthropic's 81,000-Person Survey: The People Most Exposed to AI Are the Most Worried About It

Anthropic published results from what it calls the largest and most multilingual qualitative survey of AI users ever conducted, covering 81,000 respondents across multiple countries. The headline finding won't surprise anyone who's been paying attention: people whose jobs are most exposed to AI capabilities report the highest levels of displacement anxiety. The more you use AI and understand what it can do, the more you worry about what it means for your career.

The survey found that gaining new capabilities (not speed) ranks as the top productivity benefit users report. That's a subtle but important distinction. The narrative around AI productivity has been dominated by "do the same thing faster," but users are saying the bigger value is "do things I couldn't do before." That reframes the economic impact from efficiency gains (which tend to reduce headcount) toward capability expansion (which can create new roles).

Creative professionals stand out in the data. They report feeling simultaneously limited by current AI tools and threatened by where those tools are headed. It's a uniquely uncomfortable position: the AI isn't good enough to replace you today, but it's improving fast enough that you can see the trajectory. That anxiety is rational, and it comes from the people with the most hands-on experience, not from pundits speculating from the outside.

Anthropic publishing this survey is interesting in itself. The company is essentially documenting the displacement anxiety its own products contribute to, which is either admirable transparency or a calculated move to shape the policy conversation before someone else does. Probably both. Either way, the data is valuable: 81,000 data points from actual users is a better foundation for economic policy than the usual mix of anecdotes and vibes.

Sources: Anthropic Research, Anthropic, The Decoder, Firstpost

On the Editor's Desk

ChatGPT for Clinicians reportedly outperformed doctors on OpenAI's own HealthBench Professional benchmark. We're holding this story until independent researchers can validate the methodology. Self-created benchmarks for medical claims need external verification before we report them as fact.

Meta confirmed another round of layoffs, cutting roughly 10% of staff across multiple divisions. This is the company's third major reduction since 2022, and it's happening while Meta continues to invest billions in AI infrastructure. The "cut people, buy GPUs" math is becoming a pattern across Big Tech.

Bret Taylor's Sierra acquired Fragment, a developer tools startup, signaling that the AI customer service company is building out its technical stack. Taylor, the former Salesforce co-CEO, continues to attract attention in the enterprise AI space.