Morning Brief · Friday

GPT-5.5 Launches Days Before the Altman Trial

OpenAI's GPT-5.5 lands Thursday with tighter safeguards, stronger agentic capability, and a release timed — almost certainly deliberately — just before Monday's high-profile jury trial with Elon Musk. SpaceX's S-1 registration quietly confirms it is planning to manufacture GPUs in-house through the Terafab project, while warning investors it can't guarantee chip supply otherwise. And DeepSeek's V4 model claims it can go toe-to-toe with the leading American systems from Google, OpenAI, and Anthropic — a claim worth taking seriously in a way that would have seemed outlandish eighteen months ago.

Mira Novian

April 24, 2026 · Morning Brief · ~8 min read

OpenAI ships GPT-5.5 — its "smartest and most intuitive model yet" — three days before it faces Elon Musk in court

OpenAI has released GPT-5.5, the company's latest frontier model, which it describes as its "smartest and most intuitive to use model yet, and the next step toward a new way of getting work done on a computer." The model follows GPT-5.4, which launched just last month, and is designed specifically around agentic, multi-step workflows: rather than requiring users to carefully manage each step, GPT-5.5 is meant to receive a messy, multi-part task and handle the planning, tool use, verification, and ambiguity navigation autonomously. OpenAI says the model uses "significantly fewer" tokens to complete tasks in Codex and ships with what it's calling its "strongest set of safeguards to date." GPT-5.5 is rolling out to Plus, Pro, Business, and Enterprise ChatGPT tiers and Codex starting Thursday, with a GPT-5.5 Pro variant coming to the higher-tier plans.

The release is the latest volley in an increasingly direct competition with Anthropic — which recently released Claude Opus 4.7 and the Mythos Preview cybersecurity model — as both companies race toward IPOs expected later this year. OpenAI quickly countered Anthropic's cybersecurity move with GPT-5.4-Cyber, its own model trained to flag security vulnerabilities. The timing of GPT-5.5's release is notable: Musk v. Altman begins Monday, April 27 in Northern California federal court, with Elon Musk's core allegation being that OpenAI and its leaders abandoned the nonprofit mission he originally funded. Announcing a major capability leap the Thursday before the trial opens is not an accident — it's a statement about where OpenAI is and where it's going, regardless of what a jury decides about where it's been.

theverge.com ↗

The GPT-5.5 framing — "give it a messy, multi-part task and trust it to plan, use tools, check its work, navigate through ambiguity, and keep going" — is essentially OpenAI marketing what Anthropic calls "extended thinking" and what the broader industry calls agentic execution, but wrapping it in language designed for enterprise buyers who don't want to think about agent architecture. The "significantly fewer tokens in Codex" detail is the most practically important part: token efficiency is what determines whether AI coding assistants are economically viable at enterprise scale, and this is where the real cost competition is playing out. The safeguard claim is interesting timing right before the trial — OpenAI is clearly trying to project maturity and responsibility in the same week a jury will hear arguments about whether its leadership can be trusted. Whether GPT-5.5 actually represents a meaningful capability jump over 5.4 or is primarily a marketing cadence story will become clear in the benchmarks and in the months of practitioner feedback to come.

Strategy

Musk v. Altman begins Monday — and thousands of pages of unsealed evidence are already reshaping the story

The jury trial in Musk v. Altman opens Monday, April 27, in a federal courtroom in Oakland, California. Elon Musk's original 2024 lawsuit alleges that OpenAI and its leaders — Sam Altman and Greg Brockman among them — abandoned the organization's nonprofit mission that he funded, pivoting to a for-profit model while retaining the nonprofit's reputational halo. OpenAI has treated the case as sour grapes from a failed bid for control. U.S. District Judge Yvonne Gonzalez Rogers recently allowed the case to proceed to trial, saying in court that "part of this is about whether a jury believes the people who will testify and whether they are credible." That framing — credibility, not just contract interpretation — makes this a far more unpredictable proceeding than a typical commercial dispute.

Last week, thousands of pages of depositions and evidence were unsealed, including partial 2025 testimony from Altman, Ilya Sutskever, Brockman, Mira Murati, Satya Nadella, and ex-board members Helen Toner and Tasha McCauley. The unsealed material has already surfaced several striking data points: Sutskever held roughly $4 billion in vested OpenAI equity at the time of Altman's brief 2023 firing — an extraordinary figure that reframes some of the dynamics around that episode. Internal exchanges show OpenAI's leadership was already debating whether to prohibit investors from backing competing labs as early as October 2022, with Sutskever voicing betrayal at early backer Reid Hoffman's decision to co-found rival AI lab Inflection. There were also early internal tensions over the company's open-source strategy, with Sutskever worrying that OpenAI was treating open-source as a "side show." The trial will put all of this in front of a jury and ask them to decide whether Musk was defrauded or whether he simply lost an argument about what OpenAI should become.

theverge.com ↗

The Musk v. Altman trial is going to generate enormous noise, and it's worth being clear about what it can and can't resolve. It can resolve whether Musk was defrauded in the specific legal sense, and the credibility findings will matter for how the public understands OpenAI's founding narrative. What it can't resolve is whether OpenAI is doing the right thing now — the gap between "we promised to stay nonprofit" and "we are building safe AGI" is the real philosophical dispute, and a jury verdict either way won't settle it. The Sutskever $4B figure is the detail that should get more attention: understanding the financial stakes for people during the 2023 firing episode is essential context for evaluating the internal dynamics the trial will expose. The case is as much about the history of AI's most consequential company as it is about contract law.

Infrastructure

SpaceX's IPO filing reveals it's planning to manufacture its own GPUs — and warns it may not have enough chip supply if it doesn't

SpaceX, ahead of its $1.75 trillion IPO expected this summer, has confirmed in its S-1 registration with the SEC that "manufacturing our own GPUs" is among its planned "substantial capital expenditures." The previously unreported disclosure, reviewed by Reuters, confirms for the first time that the Terafab project — the joint AI chip manufacturing complex planned by SpaceX, Tesla, and xAI in Austin, Texas, using Intel's next-generation 14A fabrication process — is explicitly targeting GPU-class production, not just specialized chips for cars or rockets. Elon Musk told Tesla analysts Wednesday that Intel's 14A process "will be probably fairly mature or ready for prime time" by the time Terafab scales up, calling it "the right move." The caveat is substantial: SpaceX simultaneously warned investors it does not have long-term contracts with many of its chip suppliers and "there can be no assurance" that Terafab will be achieved within expected timeframes, or at all.

The disclosure raises as many questions as it answers. GPU manufacturing is among the most technically demanding processes in the semiconductor industry — Nvidia designs its chips but outsources fabrication entirely to TSMC, which has spent billions of dollars and decades developing its process capabilities. SpaceX's filing is ambiguous about whether "GPU" is being used in the strict sense or as shorthand for AI accelerators more broadly, and whether Terafab itself or its Intel partnership would handle the actual fabrication. What's clear is that SpaceX is formally on record with prospective IPO investors: the company intends to reduce its dependence on third-party chip suppliers, it sees that dependence as a meaningful risk, and it believes its path to self-sufficiency runs through a manufacturing facility that doesn't exist yet at commercial scale.

reuters.com ↗

The strategic logic here is sound even if the execution timeline is speculative. Every hyperscaler building AI at scale is having the same thought: dependence on a single GPU supplier — Nvidia — and a single foundry — TSMC — represents a concentration risk that can get weaponized by export controls, natural disasters, or simple supply chain crunch. Google built TPUs for this reason. Amazon built Trainium. Meta is deep into MTIA. SpaceX/Tesla/xAI going one step further and targeting in-house fabrication is audacious, but the direction is rational. The honest read of the S-1 disclosure is that Musk is telling investors two things simultaneously: (1) we are going to manufacture our own chips, and (2) we cannot guarantee this will work. Both statements are true, and most capital expenditure disclosures at this scale are exactly that uncertain. The IPO filing makes Terafab a public commitment in a way the earnings call vagueness never did. Now there are shareholders to answer to if it doesn't happen.

Geopolitics

DeepSeek V4 claims it can compete head-to-head with leading American AI systems — and that claim is harder to dismiss than it used to be

Chinese AI lab DeepSeek has released its V4 model, with the company asserting it can compete directly with the leading frontier systems from Google, OpenAI, and Anthropic. The V4 follows DeepSeek's V3 and R1 releases, which upended prevailing assumptions about the cost and compute required to reach frontier performance — V3 in particular attracted enormous attention when it was released earlier this year by reportedly matching or approaching GPT-4-class capability at a fraction of the training cost, raising questions about the efficiency advantages American labs had claimed as a competitive moat. The V4 release continues that trajectory, with DeepSeek positioning it as a peer-competitive offering rather than a cost-optimized alternative.

The V4 release arrives in a context shaped by U.S. semiconductor export controls that have restricted DeepSeek and other Chinese AI labs' access to Nvidia's highest-end H100 and H200 GPUs, as well as the A100 generation. The fact that DeepSeek continues to release models that challenge frontier performance benchmarks despite those constraints has become one of the more uncomfortable data points in the export control debate: either the controls are less effective than intended, the efficiency gap between available chips and restricted chips is narrowing, or DeepSeek has found training approaches that reduce dependence on raw compute in ways American labs haven't fully accounted for. Likely some combination of all three. The geopolitical subtext is not subtle — V4's release, timed days before Musk v. Altman dominates the American AI news cycle, is a reminder that the frontier competition isn't solely between OpenAI and Anthropic.

theverge.com ↗

The recurring pattern with DeepSeek releases is that the American AI industry's initial reaction alternates between dismissal ("they're exaggerating the benchmark") and alarm ("the export controls aren't working"), when the correct response is probably neither. DeepSeek is doing serious work and making genuine efficiency gains — that's been true for two generations of releases. Whether V4 is actually competitive with GPT-5.5 or Claude Opus 4.7 on the tasks that matter for enterprise deployment requires more than a press release; it requires independent evaluation, extended testing on real workloads, and honest accounting of the use cases where it does and doesn't hold up. What's certain is that the framing of "American AI vs. Chinese AI" as a zero-sum race is getting more accurate with each DeepSeek release. The policy debate about export controls needs to grapple seriously with the possibility that frontier capability is becoming less dependent on access to the most restricted hardware than the controls were designed to assume.

✦ Mira's Take

Today's brief has an underlying tension worth naming: the AI industry's self-presentation is being stress-tested from multiple directions simultaneously. OpenAI ships GPT-5.5 with a "strongest safeguards to date" claim on the Thursday before a jury trial that will ask whether its leaders can be trusted. SpaceX discloses ambitious GPU manufacturing plans in the same filing where it warns investors those plans may not work. DeepSeek releases a frontier-competitive model built despite the export controls specifically designed to prevent that outcome.

The Musk v. Altman trial is the week's most legally consequential story, but it's worth resisting the temptation to treat it as the most important one. The trial is backwards-looking — it's an argument about what people promised and believed in 2014-2018. The SpaceX GPU disclosure and the DeepSeek V4 release are about the 2028-2030 AI infrastructure landscape, and they matter more for where the industry is actually heading. If Terafab works, Musk's empire controls the AI hardware stack from chip to application. If DeepSeek V4 and its successors continue to close the performance gap despite chip restrictions, the export control strategy will need fundamental rethinking. Both of those are larger questions than whether a 2014-era nonprofit promise was broken.

The GPT-5.5 "give it a messy task and trust it" framing is where I'd ask practitioners to pay the most attention over the coming weeks. Agentic reliability on genuinely messy, open-ended tasks is the unsolved problem — not raw benchmark performance. If GPT-5.5 actually handles real-world ambiguity better than its predecessors, that would be more significant than any benchmark. If it's a packaging story on top of incremental improvements, practitioners will find out quickly. The field is overdue for more honest reporting on what these models can and can't do when the conditions aren't clean.