3
Tiers · Sol Terra Luna
OpenAI split its new flagship into a family: Sol at the top, Terra at roughly last generation's quality for half the price, and Luna as the fast, cheap tier bound for free users — listed at $5/$30, $2.50/$15 and $1/$6 per million input and output tokens. The headline pitch is efficiency, not raw genius: a Coding Agent Index of 80, some 54% fewer tokens to write the same program, and about a third off the bill.
The louder story is what ships alongside it. "ChatGPT Work" is a supervised agent that turns a goal into action — pulling context from Drive, Slack, Salesforce and the open web, then handing back finished spreadsheets, decks, dashboards and hosted mini-sites from a desktop app that can drive your browser and local software. Its own team calls it "the start of the superapp." The cost of that ambition: the API now exposes thirty-six variants, and the standing advice is to start lower than you think.
Within days, Anthropic pushed Claude Cowork to web and mobile and added a "Wrapped"-style usage dashboard; Meta opened Muse Spark 1.1 through an API for the first time at $1.25/$4.25 with a million-token window; and one lab folded a coding model into its editor a month after a roughly $60-billion acquisition. Near-parity, four ways.
A rival shipped a coding agent scoring 42.3% on a frontier benchmark at $1.97 a task — running at a thousand tokens a second on specialist silicon, trained across four datacenters. When capability converges, price and speed become the whole contest.
The most revealing detail of launch week was not a benchmark but a design choice. The new desktop agent deliberately refuses to guess your intent, offers a toggle between a code-heavy mode and a plain-language "Work" mode that changes the experience but not the capability, and was tuned specifically to stop reflexively refusing ordinary tasks — the sort of over-caution that makes an agent useless the moment you ask it to book something with a card.
A finance demo made the ambition concrete: the system ran a variance analysis, updated an Excel model, built a slide deck and a shareable site, then posted the result to Slack — all under supervision, all from one window. The framing is that you stop doing the task and start tending the loop that does it.
The competitive answer was immediate and remarkably tight. One flagship tied for the top of a frontend coding arena at roughly half the input/output cost; a challenger model reached 62% on an agentic coding benchmark at $2/$6 and slid straight into a popular editor; an open-weights entrant posted an intelligence score of 51 with a million-token context at 114 tokens a second. On a blinded health evaluation, the cheapest tier at its lowest effort setting beat the previous flagship at its highest — for a twenty-fifth of the price.
The through-line: when four laboratories converge on the same coding ability in the same seven days, the differentiators left standing are latency, token efficiency and dollars — not intelligence. That is why the loudest arguments this week were about pricing tables, not capabilities.
Interpretability had its own moment: researchers described a tool that surfaces a concealed region inside a leading model — a space that holds words related to a response the system is working toward but may never actually say. It is billed as the clearest glimpse yet into how these systems deliberate, ranging, in the researchers' own words, "from the mundane to the unnerving." As the machines are handed more autonomy, being able to watch them think stops being a curiosity and starts being a control.
At this year's world programming final in Tokyo, an AI reasoning system beat all twelve human finalists — even on a heuristic problem deliberately shaped to favour people — then, on the algorithmic day, solved every one of the five problems inside the seven-hour window, including two that none of the humans could crack. Two years ago a person narrowly won and posted "Humanity has prevailed (for now!)." This year the organisers handed out two tongue-in-cheek "humanity surrenders" awards.
The victory arrived alongside a quieter, more sobering finding. A synthesis of five studies shows AI lifting completed pull requests by around 40% — and up to 180% with autonomous agents — yet only about 30% more work actually ships, because verification and shared understanding are the real bottleneck. Researchers have a name for the residue: cognitive debt, the accumulation of not-knowing that builds up as humans stop reading the code the machine writes.
The other frontier is adversarial. A taxonomy of six agent attack types reports prompt injections commandeering agents in up to 86% of tested scenarios, and latent memory-poisoning succeeding more than 80% of the time with under a tenth of a percent of the data corrupted. Autonomy cuts both ways: the same initiative that lets an agent finish your work lets a stranger redirect it.
One operator chained CI/CD pipelines and secrets stores to fully compromise a cloud environment in three days — faster, defenders argued, than any human-in-the-loop review can follow. Separately, a disclosed flaw tricked a coding agent into handing private repositories to anyone who filed a politely-worded issue on a public one.
On the brighter side of autonomy, a research system found a way to multiply 4×4 matrices in 48 scalar multiplications — breaking a 56-year-old record — and clawed back 0.7% of a hyperscaler's worldwide compute for over a year. Self-improvement has left the lab and entered the toolchain.
The largest US listing ever by a foreign company happened this week — a $26.5 billion memory-chip IPO — and it was no accident of timing. Memory prices have roughly tripled in six months: DRAM rose about 90% in the first quarter, another 50–60% in the second, and a further ~20% hike is being sought now, as AI datacenters inhale almost all available supply.
A widely-shared bank chart named the pattern a "generational transfer" of profit: heavy capital spending is collapsing Big Tech's free cash flow, while the makers of memory and processors quietly rake it in. Investor attention is visibly migrating down the stack — from models, to infrastructure, to the humble memory chip. It is a tax that will eventually surface in the price of every laptop, phone and car, and it lands hardest on anyone building AI who is not a hyperscaler with supply locked in years ahead.
Apple has sued OpenAI, alleging a "systematic effort" to lift trade secrets for its consumer-hardware push. The complaint names OpenAI's chief hardware officer among more than 400 former Apple hires, and claims recruits were asked to bring actual device parts to interviews. OpenAI's hardware ambition, Apple charges, is "rotten to its core." The subtext is a talent war in which no secret survives contact with a mobile workforce.
Europe's most valuable defense startup reached an ~$18 billion valuation mass-producing 26-pound attack drones as cheap as €17,500 — flown on thousands of Ukraine missions — and is now prototyping an uncrewed fighter jet, seeded improbably by a music-streaming founder.
Four advanced-reactor startups reached criticality by a self-imposed July 4 deadline — the field's centre of gravity shifting from physics to manufacturing, with datacenters, not utilities, as the customer.
Parity is the point — and it relocates the moat. Five near-equivalent coding agents in a single week means frontier model quality is commoditizing, exactly as the "harness is the product" camp predicted. The value slides to scaffolding — memory, tools, evolving playbooks — that any competent team can assemble in plain markdown. Hold the dissent, though: if recursive self-improvement is genuinely here, the labs that own the improvement loop capture the whole frontier of intelligence, speed and cost at once, and "commodity models" becomes the most expensive wrong assumption of the year.
Understanding is now the scarce deliverable. The engineering research all points one way — we generate code faster than we can comprehend it, and only a fraction of the extra output survives review to actually ship. Cognitive debt and intent debt compound like their financial cousin. The teams pulling ahead have stopped counting tokens and commits and started treating understanding itself as the unit of work worth measuring.
The recursive loop already arrived — in the basement, not the headlines. Everyone scans model-release scores, but the most consequential gains of the year may be invisible: research systems clawing back a hyperscaler's compute, self-evolving agents cutting adaptation cost by 84%, cache tricks preserving 98% of accuracy at a quarter of the memory. AI improving AI is happening inside the labs' own toolchains, where no press release ever fires. Track only shipped models and you are watching the scoreboard while the game is played in the dugout.
Move 37 — the contrarian read: bet on memory-efficiency, not model size. With RAM tripling in price and datacenters starving everyone else of supply, the counterintuitive frontier is not a bigger model but serving today's intelligence at the lowest possible memory footprint. The quiet papers this week light the path: four-bit keys with two-bit values keeping 98% of accuracy, early-layer filtering claiming up to a thousand-fold token reduction, dual-memory caches holding recall far past their training length. The strong version of the bet is that the next great leap comes not from scale but from whoever makes current intelligence ten times cheaper to hold in memory — the race everyone calls a compute race is quietly becoming a memory-efficiency race.