OpenAI's GPT-5.5 doubled the flagship API price over GPT-5.4 in April; OpenAI hasn't said why

Today's main line

1. [Today] Mira Murati's Thinking Machines Lab gives away its first in-house model, Inkling, and how the lab makes money is the real news

Thinking Machines Lab is the lab Mira Murati founded after leaving OpenAI, and it has raised an enormous amount of money. On July 15 it released Inkling with open weights under Apache-2.0, the most permissive commercial license. Inkling is a sparse mixture-of-experts model with about 1 trillion parameters in total and 41 billion active per token. Co-founder John Schulman announced the release (X, 2026-07-15).

Nathan Lambert, post-training lead at Ai2 and one of the most cited commentators on open models, called it the "new best American model". Lambert added that Inkling trails China's GLM 5.2 slightly on agent tasks (X, 2026-07-15).

On launch day Inkling went into Tinker, the lab's own fine-tuning service, and onto Databricks (Databricks blog, 2026-07-15). It reached the Hugging Face trending list the same day, as reported by Hugging Face's CEO (X, 2026-07-15).

For context: in June, several practitioners rated GLM-5.2 the first open-weight model that feels like a frontier model in daily use (Latent Space, 2026-06; Interconnects, 2026-06). GLM-5.2 caught up on coding and agent tasks. On general-purpose tests the gap is still wide.

Not proven yet: Inkling's specs come from Lambert's relay of the model card, not the card itself

Verification: The release itself is solid. The company, two independent commentators and the platforms' adoption all line up on the same day. Every spec figure comes from Lambert's relay of the model card. The official model card hasn't been checked directly. "Best American model" is Lambert's judgment, not a leaderboard result.

Judgment update: Who pays for open-weight models has been an open question. Our read: July 15 offered an answer that can be checked, and the order runs backward from the usual assumption. The service makes money first. Tinker is said to already bring in several hundred million dollars a year. That figure is Dylan Patel's on a podcast, a single unverified source. The flagship model is opened afterward as an asset that brings customers into the paid service. The core of this service-flywheel view comes from Lambert alone (X, 2026-07-15). Connecting it to the demand side is this publication's synthesis, and it stays a working read, not settled. The test over the next one or two quarters: whether hosted-inference companies such as Fireworks and Together follow with open models of their own.

In valuing an open-model company, where the paid service's funnel sits carries more signal than how the model itself earns money. In this one case the model is a customer-acquisition asset, not a revenue line. What would prove this wrong: Thinking Machines Lab making its next flagship closed.

2. [Today] Demand side, same day: Lovable's CEO says annualized revenue reached US$500M and more of Lovable's traffic now runs on Lovable's own open-weight models

Lovable is a leading vibe-coding company: users describe an app in conversation and Lovable writes it. CEO Anton Osika said on All-In that annualized revenue reached US$500M in May 2026, 20 months after the company was founded (All-In, 2026-07-15). Osika said about 60% of users on the cheapest plan hit its usage cap and buy more. Revenue from running apps is now growing faster than revenue from building them, by Osika's account.

The more structural statement: Lovable routes a growing share of traffic to open-weight models it post-trains itself. Lovable ranks the mistakes frontier models are making right now by customer impact and runs targeted reinforcement learning on them. The motive Osika gives is using proprietary error data to cover frontier models' weak spots, not saving money.

Not proven yet: every figure is the CEO's own account, unaudited

Verification: Everything comes from the CEO's own account and none of it is audited. The US$500M first appeared on July 14 as a third-party estimate and was confirmed by Osika the next day. A contrary account stays on record: a Wix executive has said vibe-coding can't handle complex business logic. The two don't agree, and both are kept.

Judgment update: Application-layer leaders training their own open-weight models has gone from one case (Wix and its vibe-coding product Base44) to a pattern. The motive has also shifted from cost to proprietary data. Frontier price cuts won't make that kind of demand go away. Read with item 1: on the same day, the supply side and the demand side of open models each gave first-hand accounts.

3. [Evidence update] April's quiet GPT-5.5 price rise, traced today, is the first hard data point matching Epoch AI's compute-crunch model

The fact: when OpenAI launched GPT-5.5 on April 23, 2026, it set the flagship API price at US$5 per million input tokens and US$30 per million output tokens. That is exactly double GPT-5.4's US$2.50 and US$15, and the first time in the GPT-5 series that a new generation cost more. GPT-5.6, launched July 9, keeps its flagship tier at US$5 and adds cheaper tiers at US$2.50 and US$1 (OpenAI pricing page, checked 2026-07-16; comparison from apidog, 2026-06-17).

This lines up with a theory. In May, the research group Epoch AI modeled supply and demand for inference compute. In Epoch's model, supply grows 3.4x a year against demand of about 10x, so frontier tokens get more expensive and everyday users get pushed toward smaller models (Epoch AI, 2026-05-26). Frontier prices rising while the cheap tiers multiply is the shape that model predicts.

Not proven yet: OpenAI hasn't said why it raised the price

Verification: Current prices were checked directly on OpenAI's official page. That makes this the only item in this issue that is both verified against a primary document and multi-sourced. The historical baseline relies on a third-party tracking site. The limit: OpenAI has given no official reason for the increase, and a quality premium or a pricing strategy explains it just as well. A single price point can't prove a compute squeeze, so it is weak support for the theory.

Judgment update: OpenAI's pricing now has to be read on two axes. On July 15, OpenAI President Greg Brockman framed the pitch for the flagship Sol around price per task: "our models are built to provide the best price for any given task" (X, 2026-07-15). Sol is OpenAI's name for the GPT-5.6 flagship tier above; Terra and Luna are the two cheaper tiers in the same family. Brockman also posted "6x price efficiency (!!) with Sol for react/frontend dev" without saying compared with what (6x post). A generation-on-generation doubling of the unit price and "cheaper per task" can both be true, so they are kept on separate ledgers. With the flagship unit price doubling, the open-weight route (Thinking Machines Lab's Inkling) and the application-layer self-training route (Lovable) look relatively more attractive. Product teams that depend heavily on OpenAI's API face rising pressure to negotiate or switch.

4. [Today] A resignation essay lays out the terms of Google's classified Pentagon deal

Alexander Matt Turner, a former Google DeepMind researcher, published a long essay on July 15 about why he left (Alignment Forum, 2026-07-15). The essay is the first public look at the terms of the deal, signed April 27, 2026. The reporting Turner quotes says the deal allows use for "any lawful government purpose". Google must help adjust its safety filters at the government's request. The wording against mass surveillance and autonomous weapons is "should not" language, not a binding term, and it has no enforcement mechanism.

The essay also relays a Pentagon official naming eight signatories: AWS, Google, Microsoft, NVIDIA, OpenAI, Reflection, Oracle and SpaceX. The same official called the new agreements a "counterstatement" to the dispute over Anthropic's refusal to sign. Internal opposition, at its peak: a 250-person petition, more than 600 co-signers, and Jeff Dean publicly signing an amicus brief. By Turner's account, the binding concessions it won came to zero.

Not proven yet: the contract terms are relayed reporting; the original document hasn't been seen

Verification: Concentration first: more than 70% of this item's detail comes from this one essay, written from a clear position. The contract terms are reporting Turner quotes, and the original document hasn't been seen. The only other insider reaching the same conclusion is also a dissenter. As of this issue there is no official response from Google or the Pentagon.

Judgment update: Governance that relies on character and culture, putting trusted people at the table, failed systematically in this stress test. In the same period, the only things that produced real constraints were court injunctions and red lines companies wrote into their own contracts. For due diligence, contract terms and court records carry the signal; executives' character doesn't. The essay also points to a technical gap, a single source still to be checked. Labs' safety cases rely on monitoring a model's intermediate reasoning for signs of deception. Classified, isolated deployments by default have no monitor present, so the deployment setup removes the safety case's premise.

5. [Today] "Energy is AI's natural brake" and gigawatt-scale build-outs land on the same day

Former Intel CEO Pat Gelsinger said on All-In that without power, nobody buys GPUs or builds data centers, so the bubble has a physical ceiling (All-In, 2026-07-15). Gelsinger said global energy capacity grows 4% to 5% a year, which makes this a build-out measured in decades, not years.

The supply side on the same day: Crusoe announced a 1.4 GW data-center campus in Childress, Texas, its second project with Lancium, a developer of flexible grid-connected sites (DCD, 2026-07-15). Blackstone is investing US$5.34B in behind-the-meter power projects from natural-gas midstream company Williams. Behind-the-meter means the power plant sits next to the data center and skips the utility grid's queue. The projects total 2.6 GW, all dedicated to data centers (DCD, 2026-07-15).

Not proven yet: neither deal's structure or timeline is public beyond trade-press headlines

Verification: Gelsinger's capacity figure was spoken without a source. Gelsinger is an infrastructure investor with a stake in the long-cycle story. Both deals exist so far only as headline-level trade-press reports, and their structure and timing are unknown.

Judgment update: The ceiling argument and gigawatt-scale build-outs on the same day are not a contradiction to resolve. They are two sides of one thing: power is the brake, and capital is buying power around the grid. The number to watch is whether behind-the-meter deal volume keeps growing faster than new grid capacity. If it does, the brake argument should give way to capital solving the power bottleneck on its own.

6. [Trend watch] (show aired June 30; transcript read yesterday) A Taipei energy reporter on ChinaTalk says Taiwan, the chip island, can't get into the AI compute race

A Taipei energy reporter on ChinaTalk gave the scale (ChinaTalk, 2026-06-30). In northern Taiwan, even a 5 MW data center can't get a grid connection; the reporter called such sites data centers for ants, when modern AI data centers are measured in hundreds of megawatts to gigawatts. In the south, an 80 MW project was split into four phases for lack of power. TSMC alone uses about a tenth of the island's grid.

On the policy side, President Lai Ching-te announced on March 22, 2026 that the Kuosheng and Maanshan nuclear plants meet the conditions for restart, and the official reasons list AI power demand. Maanshan could restart at the end of 2028 at the earliest. A second voice came the same week: Gelsinger, citing The Wall Street Journal, said Taiwan holds less than three weeks of energy reserves and a fab needs 90 days to restart after losing power (All-In, 2026-07-15).

Not proven yet: most of the Taiwan picture rests on one show and one nuclear-power advocate

Verification: Most of the Taiwan picture comes from one episode. The guest is a public nuclear-power advocate with a bearish view of current energy policy. Gelsinger's figures are secondhand, and the original Wall Street Journal report hasn't been checked. Two shows describing the same structure count as partial corroboration, not settlement.

Judgment update (still being checked, not settled): Compute geography may be decoupling from chip geography, and Taiwan may be the first economy-level case. The evidence is concentrated in one source, so this publication deliberately keeps the call open. The checks: whether Maanshan restarts on schedule at the end of 2028, and whether a 100 MW-class AI data-center project appears in Taiwan in 2027 or 2028. If one does, the ceiling argument needs revising.

Deep dive: Coupled in training, decoupled in deployment: our call on whether the AI harness can be split from the model (July 16, 2026)

Can a model be separated from the harness around it? The deep dive finished last night takes on this debate. The harness is the layer users actually deal with in Claude Code or Codex. Our call: the coupling happens in training, and the decoupling happens in deployment. A model is tuned inside its own company's harness, so the native pairing's experience premium on hard tasks is real. Once a task has been broken down, routed and wrapped in guardrails, which model runs which step becomes a pure price-performance question.

The old management rule that products modularize only once they are good enough gets rewritten as a frontier that moves with task difficulty. For low-difficulty work like writing a weekly report, models are already parts. For a hard ticket, the integration premium is still getting thicker. Why dig now: this is the debate with the thickest evidence in our files, and both sides are testimony from the people involved. Altman says OpenAI no longer treats the model and the harness as two things; Nadella says models can be hot-swapped. Neither has conceded in four months.

What would prove this wrong (checked within 12 months): First, a verifiable enterprise case where a hard task moves to a non-native model with no drop in quality; that kills the integration-premium call. Second, Microsoft's low-price model tier launches, high-complexity work doesn't migrate, and no common interface layer becomes a standard; that downgrades the commoditization climb. Third, multi-model routing visibly spreads out Anthropic's revenue concentration; that breaks the integrator-takes-the-profit leg. Verdict date: the first checkpoint is the end of this month. Microsoft has confirmed that within weeks it will put a fine-tuned DeepSeek model into Copilot as a low-price option. The board this debate most lacks is a clean measurement of how many points quality drops when the same harness swaps models, and the public record is empty. The 56-day cloud-versus-local test in Model watch below (arXiv 2607.13080) is the closest public number so far.

Chips & semiconductors

1. Memory prices are rising, with three signals lit at once. SK Hynix shares jumped 27% on July 14 and its market value passed US$1T. Financial journalist Joe Weisenthal pointed to the background: geopolitics shuts off cheap Chinese memory, and computer prices in the US CPI are already rising (X, 2026-07-14). Supply-chain reporter Dan Nystedt added the vendor side: Powerchip, a Taiwanese DRAM maker, raised July DRAM wafer quotes 45%, and its second-quarter revenue rose 53% from a year earlier (X, 2026-07-15). Market prices, vendor quotes and downstream consumer prices all point the same way, so this isn't noise. It runs straight into the live debate over whether AI infrastructure is overbuilt. Surging spot prices are the market voting that there is no glut right now. They also raise the cost of each gigawatt of compute and add pressure on the debt-funded path:

Our standing call (made July 14, 2026, carried here): in progress and can't be settled yet. "Circular funding hides overcapacity signals" and "zero overcapacity" are rival predictions made in advance. What would prove this wrong: compute providers disclose real utilization decoupled from their backstop clauses, which kills the masking argument; rental rates and used-GPU prices crash without warning and big orders get cancelled, which kills the zero-overcapacity argument. There is no single verdict date; the next concentrated window is the early-2027 earnings season.

2. The EUV monopolist says it is studying more capacity. ASML reported second-quarter 2026 net sales of €9.3B. Its CEO described sales as strong and said ASML is studying adding EUV (extreme ultraviolet lithography) capacity over the next two years. ASML also disclosed an agreement with Intel on Panther Lake chips (DCD, 2026-07-15). EUV machines are the single bottleneck tool for advanced chipmaking, so ASML's machine output caps the world's total advanced-node volume. The monopolist opening the door to more capacity is a vote from the very top of the supply chain that AI demand will last. The limits: this is a relay of the earnings report from one source, "studying" is not a commitment, and the official report hasn't been checked.

3. The supercomputer crown goes back to a CPU-only design. Arm's official blog cited the June 23 TOP500 list of the world's supercomputers. The new number one, LineShine, is the first system to sustain more than 2 exaflops in double precision, and it is a CPU-only, Armv9-compatible design with no GPU accelerators (Arm Newsroom, 2026-07-15). After a decade of GPUs on top, the crown going back to a CPU is a signal about architecture. The limits: TOP500's HPL benchmark measures double-precision scientific computing, not the low-precision tensor math common in AI, so the signal is confined to traditional HPC and can't yet be extended to AI training or inference hardware. This is also a vendor marketing piece with its own angle. The list can be checked at top500.org. The Arm post says nothing about LineShine's vendor or country, and that gap is worth noting.

Named commentary

1. Nathan Lambert (post-training lead at Ai2, author of Interconnects), on X, July 15, 2026. Lambert publicly estimated the scale of Inkling's final reinforcement-learning stage: about 1,000 GB300s running for about a week, at about US$5M. Lambert flagged the estimate as a guess (estimate). The value of the number is its order of magnitude, not its precision. The industry never discloses what the last mile of training a frontier-class model costs. If the order of magnitude is right, that stage costs single-digit millions of dollars, far below pretraining, and the barrier is the recipe and the data, not compute. It is a single source that Lambert himself calls a guess, so it shouldn't be cited as fact. The first official disclosure, whenever it comes, is the thing to test this anchor against.

2. Peter Steinberger (author of the open-source harness OpenClaw, now at OpenAI), on X, July 15, 2026. Steinberger amplified a claim: when a human reviewer rejects a pull request because it didn't use the right framework or follow conventions, that is a failure of automation (X). In this view, gatekeeping a machine can decide belongs to tools, and human review should move up to substantive judgment. An arXiv deployment report the same day points the same way. Review comments that were accepted got turned into version-controlled rules, and the team reports zero recurrence in the error classes covered. That is one team's self-report and should be discounted (arXiv 2607.13091, 2026-07-13). With AI producing code in volume, the bottleneck is moving to review. A note: Steinberger is an evangelist for agentic coding, and the core sentence is a quote he endorsed, not his own words.

Model watch

1. [Paper, 2026-07-11] An academic paper documents a false-memory failure in AI coding tools for the first time. In Claude Code, a command killed by a timeout can spill partial output to the screen before it dies. The conversation-compaction summary can then record that output as a confirmed result, spreading a false positive across sessions (arXiv 2607.13071). The cause in one sentence: the tool treats "seen in the terminal" as "written to disk." It is a single-author case study with no incidence statistics, but the mechanism is clean. An agent's own report that a task is done is not evidence; the artifacts left behind are.

2. [Paper, 2026-07-13] One developer tested cloud API against local open weights for 56 days, and both camps' intuitions took a hit. The same production codebase and the same coding-agent work ran for 28 days on each side. With a 99.3% prompt-cache hit rate, the cloud API's effective cost fell to US$0.57 per million tokens, below the US$2.83 amortized cost of a shared local GPU. But 74.9% of the local open-weight model's commits were fix-up commits, against 45.9% for the cloud frontier model (arXiv 2607.13080). Put plainly, "local is always cheaper" doesn't hold, and the real cost is quality and developer experience. The sample is one developer, so every number carries a single-case label. Still, this is the closest public measurement to the empty board in the deep dive above, how much quality drops after a model swap. It also shows the other side of main-line item 3: nominal unit prices are rising while caching makes effective cost collapse, both at once.

3. [Paper, 2026-07-13] "Let the model think longer" is no cure-all: planning splits into two abilities, and only one responds. A new paper ran a statistical decomposition on hard planning benchmarks across several model families. It finds that planning is two independent abilities. Local reasoning, judging whether a given step is allowed, does improve with larger models and longer thinking. Global enumeration, working out whether a goal can be reached at all, barely moves (arXiv 2607.11197). Our reading: this is one more domain-boundary counterexample to the belief that inference compute buys capability, and the sharpest one, because global search is exactly the ability that should gain most from more thinking. The lineage of that belief:

Our standing call on this line (made July 14, 2026, carried here): the two coexist and split by workload. Open-ended problems buy test-time thinking; narrow, high-frequency tasks get distilled into small models. Both are growing and they don't exclude each other. What would prove this wrong: distilled small models matching frontier reasoning models on open-ended tests would collapse the thinking-premium story; distilled models breaking down often outside their narrow domain would shrink the distillation route. This paper is boundary-drawing evidence and doesn't trigger either reversal. It is one team on one benchmark, not replicated.

Product moves

1. [Business][Linked] NVIDIA brings robot computers down to the mass-market price band, with a footnote about memory prices. NVIDIA released the Jetson T3000 and T2000 modules on July 15, bringing multimodal robot inference into the mass-market price range. The same day it launched Cosmos 3 Edge, a 4-billion-parameter on-device world model for robots that NVIDIA says can adapt to a specific robot body in a day. The modules ship in the first quarter of 2027 (NVIDIA Blog, 2026-07-15). The official copy contains one candid line: "Migrating to T3000 helps reduce costs amid high memory prices." It is the same story as the first Chips item. Even a major hardware vendor's product pitch now gets written around memory prices.

2. [Linked] LangChain bets on Slack as the home for agents. CEO Harrison Chase announced that LangChain is formally leaning into Slack as an interface for agents. The pitch is bringing AI to where teams already work, without making users switch to a new app (X, 2026-07-15). A framework vendor moving toward the distribution layer is a change of position. There are no numbers, and what the product actually looks like isn't laid out. One more one-line signal: OpenAI announced GPT-Red the same day, using automated red-team attacks to find prompt-injection vulnerabilities. It was a one-line announcement, and the product's form is unclear (X, 2026-07-15).

From the archive

1. Apple's secret self-driving experiment: 8 people and 3 to 6 million parameters put a question mark over "reinforcement learning is dead" (March 2025 material, optional catch-up). While the self-driving industry was cutting reinforcement-learning teams and betting everything on imitating human driving data, an 8-person Apple team trained a small model of 3 to 6 million parameters through pure self-play, with zero human driving data. That is less than one ten-thousandth the size of mainstream large models. Training took two to three days. In the simulator the model had one accident per million miles, against roughly one per 20,000 to 100,000 miles for human drivers (Interconnects interview with Eugene Vinitsky, 2025-03). Vinitsky, an RL veteran, offered a test that still works today: domains that can be simulated cheaply, with a small gap between simulation and reality, will eventually be taken over by reinforcement learning. Looking back a year later, "RL wins wherever rewards can be verified" turned out to be the playbook of the later reasoning-model route. It was a rare advance endorsement from outside the LLM world.

Sources & accounting

The past 24 hours. Last night's 22:07 scheduled sweep ran with zero failures and brought in 50 new pieces: 43 blog posts, 3 shows, 2 Substack newsletters, 1 company filing and 1 Stratechery piece. Of those, 4 were filtered out mechanically, 11 were read that night and 35 went onto the reading list. The overnight automated pass handled high-priority material: original X posts from Greg Brockman, Peter Steinberger and Sholto Douglas, plus transcripts of two shows, All-In and ChinaTalk. During the day, 6 arXiv papers were read (1 dropped for having no signal) and 2 targeted checks were run. No new tracked sources and no one-time catch-up today. Coverage: the main line reflects signals within this range. On X, only original posts were pulled yesterday, not a full scan.

The sources we track. The judgments in this publication rest on a long-term roster of 529 named voices: 305 on X (Elon Musk, Andrej Karpathy, Greg Brockman, Nathan Lambert and others), 90 on shows (Satya Nadella, Dario Amodei, Jensen Huang and others), 46 newsletters (Dylan Patel, Ben Thompson, Ethan Mollick and others), 48 paper authors (Noam Shazeer, Percy Liang, Tri Dao and others), 48 personal blogs (Simon Willison, Chris Olah and others), 26 companies' earnings reports and calls, and 75 institutional blogs (NVIDIA, Google Research, SemiAnalysis, Hugging Face and others). This issue uses 31 outside sources in the body.

SecondSource isn't a news digest: each day it hunts the AI firehose for the insights that matter and the practitioner judgments worth tracking over time, and shows how every item was verified. The point is always which judgment got harder and who's been right, not what happened today.

— SecondSource · generated by our research system · 31 sources · Got a view? Reply and tell us

Written from the same research and judgments as the Traditional Chinese edition. Sources are linked; we distinguish original documents from reporting and mark what we could not verify.