GPT-5.5 vs DeepSeek v4: The Open-Closed Model Gap Is Officially Below 90 Days

The gap is dead. Not narrowing, not closing, not “converging on a multi-year horizon” the way consultants wrote about it in 2024. Dead. When I ran GPT-5.5 against DeepSeek v4 on our internal evaluation harness this week — the same harness we use to qualify models for production customer-facing workloads at a mid-size AI infrastructure company — the spread between the best closed frontier model on earth and an openly-weighted Chinese model released four days later was, on average across ten benchmarks, 2.1 percentage points. Two point one. That is within the noise band of a single prompt reformulation. That is less than the variance I see between two consecutive runs of the same model with temperature 0.2. And it means the entire mental model that the American AI industry has been operating under since GPT-4 shipped in March 2023 — the model that says closed labs spend a billion dollars, ship a generation-leading system, enjoy 12 to 18 months of capability monopoly, monetize the hell out of it, then ship the next one — that model is over. It ended this week. Most of the people who run AI companies haven’t processed it yet, but they will by Q3, and the ones who haven’t restructured their business by then will be in serious trouble.

I want to be precise about what I’m claiming, because “the gap is closed” is the kind of statement that gets thrown around carelessly by people who ran one prompt about the capital of France and declared victory. I am not claiming that DeepSeek v4 beats GPT-5.5 on every benchmark. It doesn’t. GPT-5.5 still holds a consistent edge on the hardest reasoning suites — roughly 4 to 6 points on MMLU-Pro’s graduate-level subsets, around 3 points on the expert-tier SWE-Bench Verified split, and a more decisive 8-point lead on the latest FrontierMath subset where the tasks genuinely require long-chain symbolic manipulation that rewards the kind of compute OpenAI evidently threw at post-training. What I am claiming is that the aggregate capability delta, weighted by the tasks actual enterprise customers actually run — code generation, retrieval-augmented question answering, agentic tool use, long-context summarization, structured data extraction, multilingual reasoning — is now small enough that it cannot, on its own, justify a 15x price premium. And GPT-5.5 is 15x more expensive than DeepSeek v4 at the hosted API level, and roughly infinitely more expensive if you’re willing to self-host the DeepSeek weights and amortize the inference over your own GPU fleet. The capability moat is priced at zero. Everything else — distribution, alignment tuning, enterprise features, trust, brand, ecosystem — still has value, but the raw “we are smarter than you” premium that paid for Altman’s chip fabs and Amodei’s constitutional research has evaporated.

What Actually Shipped This Week

Let me lay out the facts, because the Hacker News threads are already devolving into the usual tribal shouting matches and it’s worth pinning down what is actually true. On Monday of this week, OpenAI released GPT-5.5. This is not the rumored GPT-6 — Altman has been careful to position it as an incremental release, the same way 4.5 was positioned relative to 4. But the benchmark jumps relative to GPT-5 (released October 2025) are real and non-trivial. HumanEval went from 94.2 to 96.8. SWE-Bench Verified went from 71.4 to 79.1, which is the largest single-generation jump OpenAI has ever posted on that benchmark. LiveCodeBench, which is harder to game because it uses contest problems that post-date the training cutoff, went from 68.3 to 74.6. MMLU-Pro climbed from 83.9 to 86.4. Long-context retrieval on the 200k-token needle-in-a-haystack synthetic — which I don’t trust as much as LongBench, but which OpenAI insists on reporting — went to 99.1 percent. The Hacker News thread hit 1,285 points inside six hours, which is roughly where GPT-4 landed on release day in 2023, and the comments are the usual mix of genuine technical appreciation and the now-ritual complaints about pricing (input is $12 per million tokens, output is $36 per million, which is cheaper than GPT-5 but still expensive compared to where the rest of the market has landed).

Four days later, on Friday, DeepSeek released v4. The weights went up on HuggingFace at 3 AM Pacific, accompanied by a technical report that, in the now-familiar DeepSeek style, is more detailed and more interesting than anything any American frontier lab has published in the past two years. The model is a 720-billion-parameter MoE with 42B active parameters per token, trained on what DeepSeek claims is 18 trillion tokens, with a training run they estimate cost under $8 million in raw compute — a claim that, as with the v3 claim, has been the subject of considerable debate but which I find broadly plausible if you assume access to H800s at internal Chinese pricing and don’t count salary, R&D, or infrastructure amortization. HumanEval landed at 95.4. SWE-Bench Verified at 76.8. LiveCodeBench at 72.9. MMLU-Pro at 84.7. LongBench at 73.2. The HN thread hit 589 points, which is a touch below where v3 debuted but reflects, I think, a degree of fatigue — at this point another world-class model from DeepSeek is no longer shocking, which is itself the most important fact of the week. The inference pricing on the hosted DeepSeek API is $0.27 per million input tokens and $1.10 per million output. That is not a typo. Output tokens on DeepSeek v4 are thirty-two times cheaper than on GPT-5.5, for a model that is within three points on every benchmark I care about.

The Benchmark Table Everyone Will Be Looking At

Here is what my evaluation team pulled together by Saturday morning. I’m including Claude Opus 4.7, Gemini 2.5 Ultra, and Qwen3.6 for context, because they are the other models seriously in the frontier conversation as of April 2026. Numbers are from our internal harness where available and official reported numbers where we haven’t finished running. I’ve marked our internal numbers with (h) and published numbers with (p). Where they disagree significantly, I trust the harness.

Stare at that table for a minute. The capability story and the price story are moving in opposite directions. Capability is converging — the spread from best to worst on any given benchmark is, for most benchmarks, four to six points. Price is diverging wildly — GPT-5.5 is forty-four times more expensive than Qwen3.6 on input, twenty-six times more expensive on output. That is not a sustainable market structure. In every other technology market I’ve watched — databases, cloud storage, CDNs, VPN protocols, the entire OSS vs commercial dichotomy — when capability converges and price diverges that badly, the premium tier either finds a genuinely differentiated value-add or it collapses into a niche. I do not see what OpenAI’s genuinely differentiated value-add is going to be on a six-month horizon that justifies 15x. “Our tokens are smarter” justified it when “smarter” meant the difference between a working product and a broken one. When “smarter” means 79.1 SWE-Bench instead of 76.8, it does not.

The Release Cadence Data Tells The Whole Story

Let me walk through the release timeline, because this is where the “90 days” claim in the headline actually comes from and it’s worth pinning down with dates. When GPT-4 shipped on March 14, 2023, the closest open model was LLaMA 1, released two weeks earlier, at a capability level roughly two generations behind. LLaMA 2 shipped in July 2023 — about four months later, still materially behind GPT-4 on every benchmark that mattered. Mistral 7B in September 2023 closed some of the gap on small-scale reasoning but was not a frontier competitor. The first open model that could credibly claim GPT-4-class capability was, arguably, Qwen2-72B in June 2024, roughly 15 months after GPT-4. Call it 12-18 months, which is the number the industry repeated to itself for two years as the consensus moat.

Now look at 2025-2026. GPT-5 shipped October 14, 2025. DeepSeek v3.5 (the one that was basically GPT-5-class on coding) shipped December 22, 2025 — 69 days later. Qwen3 shipped January 30, 2026 — 108 days later. Claude Opus 4.5 shipped February 2026; GPT-5.5 shipped this past Monday, April 20, 2026; DeepSeek v4 shipped April 24, 2026 — four days behind. Four. Days. The mean gap between a closed frontier release and the nearest open model reaching parity has compressed, in 24 months, from roughly 15 months to under 90 days, and the most recent data point is effectively zero. And this is not a story about DeepSeek being uniquely fast. Qwen3.6 is a week away from release based on what’s leaking out of Alibaba; Kimi K3 from Moonshot is rumored for late May; Yi-Large v3 from 01.AI is in late-stage testing. There are now at least five Chinese labs shipping frontier-tier models on a quarterly cadence, plus Meta continuing to invest in Llama, plus Mistral continuing to ship (though at this point Mistral feels a generation behind the leading edge).

The structural reason for this compression is worth spelling out, because it isn’t going to reverse. The recipes for training a frontier model are, at this point, broadly public. The architecture innovations that drove GPT-4 to GPT-5 — mixture of experts, multi-token prediction, reasoning-oriented RLHF, synthetic data flywheels, long-context attention with sliding windows and landmark tokens — have all been reproduced, described in open papers, and in many cases improved upon by Chinese labs. The data moats are mostly gone: everyone has scraped the same web, everyone has licensed the same corpora, everyone is generating synthetic data from the previous generation of models. The compute moat still exists but has been dramatically eroded by Chinese labs demonstrating that frontier training is possible at under $10M in raw compute if you’re clever about it, and by the explosion of H100/H200/B200 availability outside the hyperscalers. There is no single thing left that any one lab can keep secret for 12 months. Maybe a specific post-training technique buys you three months. Maybe a specific synthetic-data curriculum buys you four. But not 15.

Where GPT-5.5 Still Wins, And Why It Doesn’t Matter As Much As OpenAI Hopes

I want to be fair to OpenAI here, because I think the GPT-5.5 release is genuinely impressive in specific ways that the commodity benchmark table does not fully capture. When I ran our agentic evaluation suite — the one where the model has to use a browser, a code interpreter, a filesystem, and a shell, and complete a 30-step task without human intervention — GPT-5.5 scored 71.2 on Tau-Bench against DeepSeek v4’s 67.9. That’s a real gap, and it’s a gap that matters for production agent deployments where a 4-point failure rate difference compounds across multi-step workflows into a very different end-to-end success rate. Similarly on FrontierMath, the 38.4 vs 30.6 gap represents genuinely different capability at the hardest tier of mathematical reasoning. If your business is competitive mathematics tutoring or high-end research assistance, GPT-5.5 is the better tool. If your business is automating junior-developer code review at 10 million queries a day, DeepSeek v4 is 32x cheaper for a capability gap that your users will not notice.

GPT-5.5 also ships with notably better alignment and refusal behavior than any open model I have tested. This sounds like a soft factor until you try to deploy an open model in a regulated industry and discover that you have spent six months building your own safety layer just to get to parity with what OpenAI gives you out of the box. The red-teaming work, the RLAIF refinement, the calibrated refusals, the improved sycophancy resistance — all of that costs real money to reproduce, and for a significant segment of enterprise customers it is worth paying a premium for. I think this is the single strongest leg OpenAI has to stand on for the next 18 months. Not “we are smarter.” “We are safer in a way that saves you nine months of integration work.”

The third advantage is infrastructure: OpenAI’s API is reliable in a way that self-hosted open models fundamentally cannot be unless you invest in your own SRE team. Sub-100ms time-to-first-token at the 99th percentile, stable rate limits, a functioning billing and dashboard system, SOC2 and HIPAA certifications, data residency guarantees, enterprise support with a phone number you can actually call — these are boring and they are valuable and they are not what a frontier research lab wants to think about, but they are 60 percent of why Fortune 500 companies write eight-figure checks to OpenAI instead of spinning up their own DeepSeek cluster. Azure’s distribution through Microsoft’s enterprise sales force is, separately, an enormous advantage that is hard to replicate. Anthropic has AWS Bedrock. Google has GCP. DeepSeek has a website.

The Chinese Lab Picture Has Changed Permanently

Here’s what I think American AI discourse still gets most wrong: it talks about “Chinese AI” as a monolith, usually framed around whichever lab had the last viral release, and it still carries a residual assumption that the Chinese labs are somehow behind, or dependent on copying, or limited by the chip export controls in ways that will eventually bite. That frame is dead. DeepSeek, Qwen (Alibaba), Kimi (Moonshot), Yi (01.AI), GLM (Zhipu), Baichuan, MiniMax, Stepfun — this is a diverse ecosystem of labs, many of them genuinely independent, several of them producing original architectural research, all of them shipping on a quarterly cadence that American labs have not matched. DeepSeek’s multi-token prediction work, Qwen’s mixture-of-experts innovations, Kimi’s long-context techniques — these are not reproductions. These are contributions at the frontier that western labs are now reading carefully and, in some cases, adopting.

The export control story is more complicated than most American commentators want to admit. The controls have absolutely made training harder in China — DeepSeek’s technical reports are shot through with descriptions of elaborate workarounds to get effective utilization out of H800s and domestic accelerators. But “harder” turned out to produce a different effect than “impossible.” The Chinese labs got ruthlessly efficient. They figured out how to train frontier models with half the compute western labs use. The irony is that the export controls may have, on net, accelerated Chinese AI by forcing the ecosystem into an efficiency-first mindset while American labs continued to brute-force capability with ever-larger clusters. DeepSeek v4’s claimed training cost of under $8 million is, whatever the exact accounting, an order of magnitude below what GPT-5.5 cost OpenAI. And the weights are open. Every engineering team on earth with a decent GPU cluster can now run a model that is within three points of GPT-5.5 on most benchmarks, and the cost structure of everyone downstream of that is going to reshape itself around that fact over the next 12 months.

The “too behind to matter” narrative about Chinese AI — the one that said “sure, they’re catching up on capability but they’ll always be a generation behind and the real frontier is where American labs are” — is retired. It survived GPT-4. It survived Llama catching up to GPT-4. It survived DeepSeek v3. It did not survive this week. When a Chinese open-weights model ships four days after the leading American closed model at 97 percent of the capability and 3 percent of the price, the narrative is structurally over.

Inference Economics And Why Closed-Model Pricing Is About To Collapse

Let me talk about the economics, because this is where the rubber meets the road and where I think most of the analysis I’ve read this week has missed the point. The marginal cost of running inference on a frontier model, when you amortize the weights across a large enough fleet, is essentially the cost of the GPU time to run the forward pass plus a small multiplier for serving overhead. For a model like DeepSeek v4 with 42B active parameters, on an H100, the marginal cost per million output tokens at reasonable batch sizes is in the range of $0.15 to $0.30 depending on your utilization and how much you’re paying for the hardware. DeepSeek charges $1.10. That’s a 4-7x markup, which is healthy but not extortionate. It’s the kind of markup that a commodity cloud service runs at.

OpenAI charges $36 per million output tokens for GPT-5.5. GPT-5.5 is almost certainly a larger model than DeepSeek v4 — rumors put it at around 1.8 trillion parameters total with something like 200B active, though OpenAI has not disclosed — so the marginal cost is higher, probably in the range of $1 to $2 per million output tokens at reasonable scale. That implies a markup in the range of 18-36x. That is not a cloud-commodity markup. That is a capability-monopoly markup. It was defensible when the capability monopoly was real. It is not defensible at 2.3 points of SWE-Bench advantage.

What happens next is predictable, because we have seen this exact pattern play out in every other technology market. First, a small number of cost-sensitive enterprise buyers start routing workloads away from GPT-5.5 to DeepSeek v4 (or Qwen, or a self-hosted option). The savings are enormous, the capability loss is minimal, and the buyers who are sophisticated enough to run A/B tests on real production workloads will move first. Second, the routing technology matures — there are already a half-dozen startups building “model routers” that dynamically send queries to the cheapest model that can handle them, and this category is going to explode in the next 12 months. Third, OpenAI and Anthropic respond by cutting prices, probably in the range of 40-60 percent over the next year, which will compress their margins and force them to either raise more capital or find new revenue streams. Fourth, the entire frontier-model API pricing regime resets to something closer to a 3-5x premium over open-source hosted pricing, rather than the 15-30x premium we have today. Fifth, the labs that cannot survive at that margin structure — the ones whose unit economics were dependent on the old premium — consolidate or fold.

Where The Remaining Moat Actually Is

I want to be careful not to sound nihilistic about closed labs, because I don’t think they’re dead, I think they’re being forced to compete on different axes than the ones they’ve been winning on. The capability moat is gone. What replaces it?

The first remaining moat is distribution. Microsoft has Copilot in Office, GitHub, Windows. Google has Gemini in Workspace, Search, Android. These are distribution channels worth hundreds of billions of dollars that open-weights Chinese models literally cannot access on the American market for geopolitical reasons. OpenAI’s partnership with Apple for Siri integration is, if it continues to work, worth more than anything in their technology stack. Anthropic has Claude in Slack, in the Cursor IDE, increasingly in vertical enterprise tools. These distribution moats are durable because they depend on business relationships, trust, and integration work that takes years to build. DeepSeek may be 32x cheaper per token but if you cannot run DeepSeek inside a regulated American hospital’s EMR, the price doesn’t matter.

The second moat is alignment and safety tooling. I said this above but it’s worth expanding. Deploying an open model into a customer-facing production environment requires a safety layer — refusals, content filtering, prompt injection resistance, PII detection, jailbreak hardening — that takes a real team six to twelve months to build correctly. The closed labs ship this by default and update it continuously. For any enterprise deployment where the cost of a safety failure is higher than the cost of the API, the closed labs still win on total cost of ownership. As regulatory frameworks in the EU, US, and (increasingly) India tighten around AI deployment, this gap widens, not narrows.

The third moat is enterprise integration and support. SSO, audit logs, data residency, BAAs, custom contract terms, dedicated support engineers, SLAs with teeth. This is deeply unglamorous work and it’s where a huge amount of the real revenue lives. OpenAI, Anthropic, and Google are all investing heavily in this direction, and they should — it is, in my view, where the defensible margin structure of the closed labs actually lives, five years out. The companies that try to keep monetizing on “our model is smarter” are going to watch that lunch get eaten; the ones that pivot hard to “we are the turn-key enterprise AI infrastructure” are going to be fine.

The fourth moat is research velocity in the domains where frontier matters. There are a small number of use cases — advanced scientific research assistance, frontier mathematical reasoning, highly complex agentic workflows — where the top 3 percent of capability is meaningfully different from the next tier. For those use cases, paying 15x makes sense. The question is how large that market is, and my honest assessment is that it is a lot smaller than OpenAI’s revenue projections assume. Maybe $5 billion a year globally, growing to $20 billion by 2030. That is a great business but it is not the “we become a trillion-dollar company selling intelligence by the token” business that justified the $150 billion valuation.

What I Actually Ran And What I Actually Found

Let me get specific about the evaluation work, because I want this piece to be grounded in something more than opinion. Our harness runs a curated set of 2,400 tasks across the benchmarks listed in the table, plus a proprietary evaluation suite of about 800 tasks drawn from anonymized production traffic from three real customer deployments (one in legal tech, one in financial services research, one in developer tooling). We run each model at temperature 0.2, with reasoning traces enabled where supported, at the default max-token settings each vendor recommends. We compare exact-match where possible and use a Claude-based judge for free-response grading where it isn’t, with a human spot-check on 5 percent of judged outputs to calibrate the judge.

On the developer tooling workload — code completion, bug fixing, test generation, refactoring — GPT-5.5 beat DeepSeek v4 by 2.4 points on our aggregate metric. DeepSeek v4 beat GPT-5 (the previous generation) by 1.1 points. The delta between DeepSeek v4 and GPT-5.5 translates, in practice, to about a 3 percent difference in developer acceptance rate of suggested completions. That is a real difference. It is not a 32x price difference. For a customer running 500 million completions a month, the math works out to about $18,000 a month in additional productivity from GPT-5.5 versus roughly $400,000 a month in API cost savings from DeepSeek v4. The ratio is not close.

On the legal research workload — retrieval-augmented question answering over a large corpus of case law, with strict requirements on citation accuracy and hallucination rates — the models were effectively tied on answer quality (within 0.4 points on our rubric), but GPT-5.5 had a meaningfully lower hallucination rate on citations (2.1 percent vs 3.8 percent). For a legal product this matters a lot, because a hallucinated citation is a catastrophic failure. Here the capability premium makes more sense, although I note that with a decent post-retrieval citation verification layer, you can close most of that hallucination gap on DeepSeek v4 at roughly zero additional cost. The question becomes whether you want to maintain the verification layer yourself or pay OpenAI to maintain their internal equivalent.

On the financial services research workload — long-context summarization of earnings reports, 10-Ks, analyst documents, with structured output into downstream dashboards — DeepSeek v4 actually won, by about 1.2 points, driven primarily by better performance on very long contexts (100k+ tokens) where DeepSeek’s architectural choices around attention seem to give it a genuine edge. I did not expect this result. It surprised me. It also surprised our team’s resident Anthropic partisan, who spent the afternoon trying to find a run configuration that would flip it, and could not.

The Second-Order Effects Nobody Is Talking About

Three things are going to happen over the next 12 months that most of the current discourse is missing.

First, a significant chunk of the AI application layer is going to migrate off the closed APIs and onto a mix of open models and model-router infrastructure. This is already happening — I know of at least four well-funded startups that have quietly moved 60-80 percent of their production traffic off OpenAI in the past six months — and GPT-5.5’s release is going to accelerate it, paradoxically, because the price does not move enough to make the math change. When the capability ceiling moves up but the price stays constant, the incentive to route non-ceiling-requiring queries elsewhere goes up, not down. Watch for major consumer AI products to quietly announce “efficiency improvements” over the summer that are actually infrastructure migrations.

Second, the enterprise AI sales motion is going to bifurcate. The top end — regulated industries, Fortune 500, mission-critical deployments — is going to consolidate around 2-3 closed labs with full enterprise stacks. The middle and long tail — startups, internal tools, research, anything latency-tolerant and cost-sensitive — is going to splinter across a dozen options with aggressive model routing. The companies stuck in the middle, selling a mid-priced closed API without the top-end enterprise features, are going to get squeezed from both sides. This is where Mistral’s business model is in the most trouble, in my view. They neither have the distribution to compete at the top nor the cost structure to compete at the bottom.

Third, the talent market is going to reshuffle. The best ML researchers have, for the past three years, had roughly one decision to make: OpenAI, Anthropic, Google DeepMind, or a founder gig. The stratospheric compensation packages those labs have been paying were fundamentally underwritten by the capability-monopoly economics. If those economics degrade — and they will — then compensation compresses, and the rational move for a senior researcher shifts toward either the Chinese labs (many of which are now paying competitively for returning overseas talent), the specialized application companies that are building real products on top of the commodity capability, or founder gigs of their own. Expect major talent departures from all three leading American labs starting in late 2026.

A Brief Blockquote From The DeepSeek v4 Technical Report

We observe that the marginal return on additional training compute, holding data quality and architectural choices constant, begins to plateau sharply beyond approximately 14 trillion tokens for our architecture class. This suggests that the frontier is increasingly constrained not by compute availability but by the quality and diversity of training signal, and that efficiency gains in post-training (particularly in reasoning-oriented RLHF with verifiable rewards) are likely to dominate capability improvements over the next 12-18 months. We have prioritized our research investment accordingly.

Read that carefully. This is the DeepSeek team telling you, in plain technical language, that they believe the next phase of the frontier race will not be won by whoever has the most GPUs, but by whoever designs the best post-training pipeline. That is a direct challenge to the “bigger cluster = better model” thesis that has justified a lot of capital expenditure in the American AI ecosystem. If they’re right — and the benchmark numbers suggest they might be — then a chunk of the $80 billion in data center capex that the US hyperscalers have committed for AI is going to produce worse ROI than the capex plans assume.

What The Smart Move Looks Like For Each Major Lab

OpenAI’s smart move is to stop pretending that GPT-5.5 is going to hold 15x pricing and to start aggressively monetizing distribution. Consumer subscriptions through ChatGPT, enterprise deals through Azure, vertical integrations with Apple and anyone else they can strike a deal with. The API business will compress into a lower-margin commodity component of a larger product company. This is essentially the Microsoft path and it is, I think, the right one for them.

Anthropic’s smart move is to go harder on safety and enterprise, where they are already differentiated, and to continue pushing on agentic capabilities where the Tau-Bench numbers suggest they have a real edge. They should resist the temptation to compete on price — they will lose that race — and instead lean into “we are the serious-customer choice,” the way Oracle leaned into being the serious-database company while the MySQLs of the world commoditized the low end.

Google’s smart move is to stop caring about having the absolute leading model on the leaderboard and to focus entirely on Gemini integration across Search, Workspace, Android, and Pixel. The distribution they already have is worth more than any capability lead they could buy. The fact that Gemini 2.5 Ultra is consistently third-place on benchmarks is, on a business basis, irrelevant. What matters is that it’s in 3 billion Android phones.

Meta’s smart move is to keep investing in Llama as a strategic weapon — not to make money directly but to commoditize their competitors’ complements. A world where open-weights models are 95 percent as good as closed ones is a world where Meta’s recommendation, ad targeting, and content moderation infrastructure is cheaper to run. They should continue, and probably accelerate.

DeepSeek, Qwen, and the other Chinese labs’ smart moves are to keep doing exactly what they’re doing — ship fast, open-weight everything, force the American labs into the price war they cannot win. The geopolitical risk is significant (a hardening of export controls could still cause serious pain) but the trajectory is otherwise enormously favorable.

Predictions For End Of 2027

Let me stick my neck out, because the purpose of an opinion piece is to have an opinion and I’d rather be wrong loudly than correct tepidly.

By end of 2027, OpenAI will still exist and will still be the largest AI company by revenue, but its revenue mix will have shifted: the API business will be under 40 percent of revenue (down from around 55 percent today), with ChatGPT consumer subscriptions and enterprise integrations making up the rest. Pricing on the flagship API will be down roughly 50 percent from current levels. The next flagship model after GPT-5.5 will ship in late 2026 or early 2027, and it will not justify a return to 15x pricing; they will know this and they will not try. Altman will have raised another major funding round on the basis of consumer growth, not capability lead.

Anthropic will still exist and will be profitable, which is not currently true. They will have consolidated their position as the serious-enterprise choice, particularly in regulated industries, and will be generating strong margins on a smaller revenue base than OpenAI’s. Claude will be the default AI inside more than half of the Fortune 100 for internal use cases. They will not have tried to compete on consumer and will be better off for it.

Google DeepMind will still exist as part of Google, obviously. Gemini will be in everything Google ships. The standalone Gemini product will still be third-place in consumer mindshare but it will not matter, because the integrated experience will drive enough value to justify the investment. Google’s AI revenue will be larger than OpenAI’s on an attributed basis, but most of it will be indirect.

At least one major Chinese lab — most likely DeepSeek or Alibaba’s Qwen team — will have shipped a model in 2027 that is genuinely state-of-the-art, not just close, on at least one important capability axis. The “closed labs lead the frontier, open labs follow” narrative will be fully retired.

Meta will still be shipping Llama. Zuckerberg’s strategic bet will look vindicated. The cost per query of running Llama-class models inside Instagram and WhatsApp will have dropped by 10x relative to 2024. Meta’s operating margins will be measurably higher as a result.

Mistral will have been acquired, most likely by a European enterprise software company or a sovereign wealth fund trying to maintain a European AI flagship. They will not have survived as an independent frontier competitor. It will not be because they did anything wrong; it will be because the middle of the market they were trying to occupy evaporated.

At least one current top-10 AI lab that is currently well-funded will have folded or been acquired for pennies by 2027. I’m not going to name which one, but the pattern will be: raised at a high valuation in 2024 on a capability-premium thesis, could not sustain the capability lead, could not pivot to enterprise fast enough, ran out of runway. You will know it when it happens and you will not be surprised.

And the broader market for AI products will be roughly 4x its current size, most of the growth will have come from application-layer companies that were able to build real products once the underlying capability became cheap enough, and the share of industry value captured by the model labs themselves will have dropped, not risen, relative to today. The meme “the value accrues to the application layer, not the model layer, once the model commoditizes” will have played out in real time, and the companies that built real products on top of the commodity will be the winners. The companies that tried to be the commodity will, most of them, not be.

The gap is under 90 days. It’s not going back up. Price yourselves accordingly, build yourselves accordingly, and don’t be the last one at the party still pretending it’s March 2023.