Yesterday the duopoly halved the headline: GPT-6 Sol and Luna landed at $2/$10 and $0.10/$0.50, and Opus 5.5 at $4/$20 with its cache line cut 60 percent. Hours later, OpenAI shipped the other half of that war — the part operators actually pay on. A “Better prompt caching for GPT-6” release adds a caching dashboard, a tool that names why a cache miss happened and how many tokens it cost, explicit breakpoints, and prewarming, all aimed at one number: your cache hit rate. When both labs are done cutting list prices, the player who squeezes a working cache out of every session is the one who controls the recurring bill, and this week each lab moved to make hit rates something you engineer instead of something you assume.
The fifty percent price cut only lands if your hit rate does
What happened. Within hours of OpenAI halving GPT-6 Sol and Luna list prices on Tuesday, it published “Better prompt caching for GPT-6”: GPT-6 caches shared prefixes by default with discounts up to 90 percent on cached input tokens, reused within a 30-minute window. The new tooling is where the engineering is. A Prompt Caching Dashboard tracks hit rate over time and splits input into cached and uncached tokens. A diagnostics endpoint compares a request against a recent response and names the miss reason — tools_changed, a model or settings change — with an estimated count of affected tokens. Explicit breakpoints let you choose which prefixes get cached, and you can now change reasoning effort mid-conversation without breaking cache. The guidance is pointed: keep tool definitions, schemas, and ordering stable, use allowed_tools instead of removing definitions, append instructions rather than rewriting them, and prewarm the cache to push context-processing out of user wait time. Anthropic had set the pace a day earlier, cutting Opus 5.5’s cache reads to $0.20/M, 60 percent below its own Opus 5.
Why it matters. This is the logical second act of a price war that opened with a rate card. A list-price cut is only money in the bank if you actually cash it in, and for a persistent agent the cash registers are cached reads — on long agentic sessions those are the majority of input tokens, so a few points of hit rate are worth more per agent-hour than another ten percent off the headline input price. The reason this release matters is that it turns cache economics from a vendor promise into an operator metric: a dashboard, a miss forensics tool, and explicit breakpoints are the instrumentation of a cost you used to take on faith. OpenAI is the one publishing it — hours after it launched GPT-6 Sol and Luna at half price and three hours after the Tuesday card made the same argument on Anthropic’s side: the cache line is where agents actually bill. The vendor making its cache cheap is now also making your hit rate auditable, which is exactly the transparency that turns a discount into a switching cost.
Source: openai.com
OpenAI’s defenders are now inside Ukraine’s civilian infrastructure
What happened. OpenAI announced it will give the Government of Ukraine access to its Daybreak cyber-defense program, working with the Ministry of Digital Transformation to identify software vulnerabilities and develop and test fixes faster. The announcement came on the sidelines of the UN General Assembly, from Kyiv’s consul general in San Francisco and OpenAI’s head of national security policy. The context on the primary is specific: CERT-UA handled nearly 6,000 cyber incidents in 2025 across hospital, energy, and telecom systems. OpenAI also points to prior European deployments — France, Germany, Poland — where its models helped ENISA and Poland’s CERT Polska find vulnerabilities (six in third-party router software, since patched).
Why it matters. Last week the Maven story was AI at the point of fire, targeting software that a probe pulled back from the kill chain. This is the same war, AI on the defensive side: patching civilian infrastructure before the attacks land. What the pair shows is an emerging asymmetry in how the AI industry deploys into conflict — offensive targeting drew a Pentagon probe and a rewritten command loop, while defensive vulnerability-hunting earned a government partnership announced at the UN. That line matters operationally: an AI that defends civilian infrastructure is treated as the deployable lane, one an AI vendor can open well before it would ever ship the offensive counterpart. For anyone building defense-adjacent AI, which side of that line your tool sits on is becoming a compliance question before it is an ethics one.
Source: openai.com
a16z is betting the credential is the network, not the degree
What happened. Andreessen Horowitz is putting $35 million into the Horowitz Andreessen Academy, a for-profit private academy in San Francisco whose stated product is turning high-school graduates into founders. No grades, tests, or homework; the first founding class of about 50 students is a year of tuition-free “building” in the city, with partners from Anthropic, OpenAI, Google, and Meta. A paid two-year unaccredited program is planned for fall 2028 at elite-private tuition. The details that matter are the ones that break the usual model: students own 100 percent of the intellectual property they create, and the venture firm takes no stake. Gagan Biyani, a Udemy cofounder, runs it; Garry Tan and Tobi Lütke are among the backers.
Why it matters. The no-grades marketing aside, the $35 million placement is a bet on where AI-era labor value lives. A firm that makes its living on equity investing is funding a pipeline where it takes no equity — so the return has to be the standing view of the most ambitious building cohort in the city, working with the frontier labs, while the firm watches what they make. That is the credential market being repriced in real time: not a degree that signals pedigree, but a network plus portfolio-proof that demonstrates output. For a systems engineer the read is structural, not ideological — when capital funds its own talent supply chain instead of renting it from universities, the input cost of that labor just became a firm’s own balance-sheet line.
Source: sfstandard.com
The Rest
- Stanford used AI to swap a Latino student’s face for a generated Black woman on campus banners, editing other students slimmer, and pulled the banners as a policy violation. The failure wasn’t the generator — it was the unchallenged “make it representative” instruction and the missing human sign-off on a real person’s likeness. sfstandard.com
- Israel picked state-backed funds to drive investment in next-generation defense tech, each mandated to raise at least $150 million for startups, a sovereign push straight into the defense-AI lane this column keeps watching. timesofisrael.com
- The UAE central bank sanctioned local branches of Iran’s Bank Melli — Gulf financial pressure on Iran is now a banking act, landing as US–Iran talks open at the UNGA. timesofisrael.com
- Israel’s sixth submarine is arriving as the Navy plans a major fleet expansion — a quiet maritime-deterrence signal that keeps getting capital even while the Gulf’s political track floats negotiation. timesofisrael.com
- The Netherlands ended its Eurovision boycott over Israel, announcing a return for 2027 — a small de-isolation datapoint against a year of diplomatic pressure campaigns. timesofisrael.com
What I’m watching
Whether cache tooling becomes the battleground rather than the rate card — the operator benchmark that matters now is the cacheable-prefix ceiling per vendor, and a vendor who stays opaque on hit rates loses the audit. Whether Anthropic answers with equivalent diagnostics on Sonnet 5.5 and Haiku 5.5, where the price war’s bottom end actually lives. And whether OpenAI’s next move before DevDay on September 29 is a system release or another card — the cache control-plane is a reasonable preview of what a systems team ships when models stop being the differentiator.