The Brief

今日头条

The frontier model just stopped being a paid product. As of this week the newest generation of OpenAI’s flagship — GPT-5.6, in its “Luna” variant — is the default model for free and Go-tier ChatGPT users, with unlimited text conversation, at a scale OpenAI puts at one billion weekly users. The thing that used to be the upgrade, the generation behind the paywall, is now the floor.

The more consequential part is less the model than the dial. OpenAI is shipping a thinking-intensity slider across web, mobile, and desktop that lets users decide how much inference time each answer consumes, alongside a “Think” button on the free tier for genuinely hard questions. For Plus and Pro users the same GPT-5.6 Sol model handles both instant answers and deep reasoning, so moving the slider changes time-invested, not persona — the design goal is that a user switching from a quick answer to full reasoning should feel the model working longer, not suddenly talking to a different machine.

That is the quietly big move here. Reasoning depth, for years the hidden cost variable and the differentiator between tiers, has been turned into a visible, user-facing control served to a billion people. The paid-versus-free split is no longer about which model you get; it’s about the range of the dial, plus the tool and upload limits that remain on the free tier. “Unlimited text conversation” comes with an asterisk worth noting: it is still subject to anti-abuse mechanisms, and file uploads, images, and other tools stay outside the unlimited scope.

OpenAI’s internal numbers give the update its substance. In an evaluation covering finance, medical, and legal questions requiring specific factual support, answers containing at least one factual error fell by about 62% with Luna and 68% with Sol compared with GPT-5.5 Instant. Those are self-reported, from a company that also said the updated model makes better use of search-and-retrieval sources — especially on dates, numbers, sources, rules, and assumptions, which is precisely where conversational models have been weak. Sol is also tuned to answer simple questions directly and skip padding, and to correct the user when simple agreement would be unhelpful, which is a small but real behavior change.

One boundary matters: the new Sol is optimized specifically for ChatGPT’s chat experience, and the version of Sol powering ChatGPT Work and Codex does not change with this release. So this is a consumer-interface update riding on the company’s flagship name, not a rewrite of its agentic products. On the safety side, OpenAI says the accompanying system card details training and evaluation measures, including new protections aimed at users who may be under 18.

What this adds up to is that OpenAI has collapsed the distance between the paying and non-paying experience of its flagship while making thinking effort a consumer commodity. That is a statement about where the business is heading: with a billion weekly users, the constraint is no longer model quality at the top of the funnel but the cost of serving reasoning to the bottom of it.

Watch the dial’s free-tier ceiling, and what happens to the “Think” button under load. The default position of that slider, and how aggressively the anti-abuse limits throttle heavy use, will determine whether this is generosity or a taste of a paywalled future — and it’s the decision that decides whether the subscription premium has, in effect, migrated from model access to thinking budget. OpenAI给10亿用户免费换上GPT-5.6

还有这些

Shopify replaced Redis with MySQL for inventory reservations–and it scaled · Source Shopify’s oversell-protection system — the thing that keeps two buyers from claiming the same last unit — moved from Redis to MySQL and scaled. The team rebuilt it on a one-row-per-unit design using MySQL 8’s SKIP LOCKED, a capped 1,000-row replenishment pool per item and location, a composite primary key to cut lock count, READ COMMITTED to dodge gap locks, and UNION ALL batching. The real ceiling turned out to be connections, not CPU: after tagging every SQL statement by business process and aggregating hold time at ProxySQL, they found checkout code outside reservations was squatting on the pool. Cleaning it up removed 50% of primary-database reads and 33% of transactions; the migration ran in shadow mode beside Redis until MySQL proved itself on live traffic. The durable lesson is that the bottleneck wasn’t the database engine at all, but whoever was holding connections they no longer needed.

Microsoft Edge is about to lock out older ad blockers, just like Chrome did · Source Microsoft Edge is ending Manifest V2 support, and with it the classic uBlock Origin. Microsoft says just 58 extensions on the Edge store with meaningful usage still run on MV2, and only three lack MV3 equivalents, so the phase-out begins this month in Canary, Dev and Beta, reaches Stable through the rest of 2026, and hits managed devices in early 2027. The company points holdouts to uBlock Origin Lite and notes that Opera still honors MV2 ‘for as long as it’s technically reasonable.’ Chrome already pulled the same switch a year ago, so the interesting number is how small the protest has gotten. The ad-blocker arms race was effectively over before Edge even flipped the switch.

Embattled hedge fund Situational Awareness invests $400M in chip startup Source Foundry · Source Situational Awareness, Leopold Aschenbrenner’s hedge fund, put another $400 million into the Stanford-founded chip startup Source Foundry, taking its total stake to $500 million — even as the fund’s assets collapsed from $20 billion to $10 billion in the AI-infrastructure rout. At the end of July it sold most of its public portfolio to Citadel but kept its Anthropic shares. A WSJ-reported $400 million into a stealth fab venture, made weeks after unwinding most of its book, is a remarkable display of risk appetite. It suggests Aschenbrenner’s conviction that hardware manufacturing is the bottleneck has not wavered, even when his public positions were being unwound around him.

苹果公司秘密测试中国长鑫存储芯片 应对全球存储芯片荒 · Source Apple is quietly testing DRAM from China’s CXMT as the AI-driven memory shortage forces it to raise prices globally. HP and Acer already use CXMT chips in devices sold outside the US, and CXMT — fresh off a Shanghai IPO — now takes 7% of global DRAM revenue after an eightfold year-over-year revenue jump in Q2. But the fit is poor: US export rules forbid transferring the technical specs Apple’s custom silicon requires, and insiders say CXMT’s capacity for the year is already booked out to ByteDance, Tencent and Xiaomi. Even with capacity doubling planned by 2028, the report concludes CXMT can’t be Apple’s cure. Politics aside, Apple’s usual trick of custom-partnership leverage simply doesn’t work here.

Stripe Uses Graph Search and State Machines to Automate Database Remediation · Source Stripe replaced its hard-coded database-remediation runbooks with a graph model of its MongoDB fleet, using Dijkstra’s algorithm and state machines to compute recovery plans on the fly. The old plugin system paged operators 124 times in six months for misconfigured shards and blocked index builds for an hour per incident; the new approach cut database-related pager alerts by about 30% and eliminated an estimated 12 days of unhealthy shard states a year. Because Dijkstra explores all reachable states, it returns partial remediation when no complete path exists, and Stripe plans to extend it to topology changes and blue-green deploys. The payoff is that it finds recovery paths no runbook author ever thought to write down.

快讯

• Windows 11’s built-in Weather app has been flagged for consuming more than 1 GB of RAM, another data point in Microsoft’s uneven performance cleanup. (Source)
• The UTM project’s Triton driver brings DirectX 11 to QEMU guest VMs, building on its earlier Neptune Direct3D protocol-forwarding layer. (Source)
• GitHub has retired GitHub Models, its free AI model playground, as Simon Willison noted on his blog. (Source)
• An 11-year bet from 2011 on whether a prediction’s original URL would still resolve has come due, putting Tim Berners-Lee’s ‘cool URIs don’t change’ on the line. (Source)
• Hackaday’s review of Intel’s latest laptop silicon asks whether the company has finally closed the performance-per-watt gap with Apple Silicon. (Source)
• Two Georgia police officers were fired over misuse of Flock’s license-plate-reading cameras, widening the surveillance vendor’s municipal fallout. (Source)
• 49ers head coach Kyle Shanahan says his Tesla was on Autopilot when he crashed, adding another high-profile incident to scrutiny of the system. (Source)
• David Ellison won backing from the world’s two largest movie theater chains for a merger their peers oppose, a key step for his Paramount play. (Source)
• AI-detection software is producing a new era of distrust, flagging legitimate human writing as machine-generated and eroding credibility in classrooms and newsrooms. (Source)
• Apple is exploring a round Apple Watch, a screenless fitness tracker and other radical redesigns, per Bloomberg’s Mark Gurman. (Source)
• A player broke out of Nvidia’s GeForce NOW sandbox to reach a full Windows desktop, exposing the limits of the cloud gaming service’s lockdown. (Source)
• A Carnegie Mellon study found the productivity gain from AI-written code fades after about three months while static-analysis warnings accumulate, a case for guardrail tooling like Sonar’s. (Source)

一句话

Runbooks encode known recovery procedures; a state machine discovers novel ones.

— Stripe engineering team, on their graph-based database remediation system (via InfoQ)