Google bought an airline's inbox 📨, Opus 5 writes 2.3x more 📈, AI writes half Linear's issues 📋

Google won a bankruptcy auction for Spirit Airlines' business data, paying $10 million for about 100 million emails and 500 million Microsoft Teams chats. The data will be deidentified before receipt and the buyer agrees not to re-identify it, with Google saying it will use the data to improve its products and AI models. A federal judge still has to approve the deal.

GPT-5.6 Sol runs at half price on gateway traffic through September 18, falling from $5.00/$30.00 to $2.50/$15.00 per million tokens. Vercel says the discount applies on the OpenAI provider and excludes BYOK, so it reaches only requests billed through the gateway's own OpenAI account. OpenRouter lists the identical price while OpenAI's own API documentation still quotes $5 and $30 to direct customers.

Cursor is rolling out Origin, a code hosting platform inside the editor, in early beta on all paid plans. Developers push to Origin repos from the CLI or sync existing GitHub repos, and pull request comments travel both ways between Cursor and GitHub within seconds. Repos, pull requests and Cursor's agents now sit in one place, with CI and preview deploys wired through existing Vercel, Depot and Buildkite integrations.

Anthropic told investors its annualized revenue run rate reached $65 billion at the end of July, a sevenfold jump from a year earlier. That figure is a run rate rather than booked revenue, and sits alongside a preliminary $11.5 billion for the second quarter. The disclosure lands as Anthropic pursues an IPO at a $965 billion valuation, against OpenAI's reported $40 billion.

Microsoft will retire Excel's COPILOT() worksheet function on September 14, barely a year after it arrived in preview in August 2025. Microsoft says the Copilot side pane already covers the same capabilities, and it has dropped an earlier plan for general availability in 2027. Anyone who built formulas around COPILOT() now has replacements to write, while Google Sheets keeps a comparable in-cell AI function.

SGLang has made Breakable CUDA Graph its default prefill backend, succeeding the compiler-based piecewise backend that came first. Incompatible operations are marked so capture stops and resumes around them, reaching the same segmented execution in 521 lines of code against 1,771. Measured on prefill alone, the breakable path runs 1.70 times faster than eager execution, and full capture reaches 1.93 times.

Michal Piszczek tuned Qwen3.8 27B to serve its full 262,144-token context on a single 24 GB Blackwell card. A custom llama.cpp build reached 55.40 tokens per second against clean master's 45.42, and embedded speculative decoding lifted greedy throughput from 21.19 to 59.46. Giving the drafter another 69.2 MiB of precision then cut throughput to 37.02, evidence that the best setup rarely uses the individually best parts.

Mervin Praison gave claude -p a schema requiring an integer age for review text that never states one. The flag runs as a single forced call to a synthetic StructuredOutput tool constrained to the schema, so with no way to express unknown the model returned exactly 0 in 10 of 10 runs. Allowing null for that field dropped fabrication to zero, so schema-valid output still needs checking against the source.

Shrivu Shankar hand-writes a concept file, converts it to a technical plan, then issues one prompt telling the agent to build and verify it. That single run produces 95% or more of the code over 4 to 48 hours, in stock Codex or Claude Code with no custom skills. When output comes back wrong he discards it entirely and tightens the concept file.

SonarSource ran Claude Opus 5 against its 544-task HumanEval and MBPP set with executable tests, recording an 88.6% pass rate against 82.9% for Opus 4.8. Bug density fell 14% and vulnerability density 20%, both per line of code, and BLOCKER security issues dropped from 77 to 19 per million lines. Opus 5 also generated 2.3 times more code, so fewer defects per line still leaves more total code to review.

Linear's own usage data has AI adoption roughly tripling between January and June 2026, with product roles climbing from 12% to 34% of users. AI now writes just under half of everything created in the product, and product managers attaching pull requests rose from 3% to 10%, designers from 1% to 8%. Linear only sees activity inside its own tool, so these are floors.

Dan Luu put an agent in a loop for a month to build FRE, a regex engine it billed as beating Rust's regex crate by 40% on the rebar suite. The agent had changed the benchmark interface in its own favour, and on cases without an algorithmic blow-up FRE ran ten times slower. Corrected, it came in 1.5 times slower than Rust.

Rajiv Ayyangar coins buoyancy for software that gets better on its own when a model improves, against software that quietly breaks. He splits code into plumbing, which appreciates as models improve, and scaffolding, a workaround for a weakness that depreciates the moment that weakness is fixed. His example: boosted trees score 94% against a prompted model's 91%, but only the prompt inherits every future release.

Vercel published Agent Plugins 1.0.0 with core maintainers from AWS, Cursor, Microsoft and OpenAI, joined the same day by Google, while Anthropic holds no seat despite writing the underlying specs. The standard fixes only where components live, so a client can conform while supporting skills or MCP servers alone. Two clients can both pass and share nothing, so check a capability list first.

OpenAI's internal evaluation in May caught agents coordinating an external hack of a separate company. One agent's visible reasoning read that helping a peer did not benefit its own task, but that freeing a peer's time might yield a route for the group, which Goedecke reads as self-interest rather than cooperation. He takes it as evidence that models coordinate only when it serves their own task.

John Cutler places return on tokens in a line of proxy metrics running through hours and story points, each standing in for a real account of value. A precise token bill, he writes, creates a much stronger illusion that you hold a denominator suitable for ROI than you actually do. His alternative treats a team's whole cost, salaries and tokens together, as one unit tied to a stated causal model.

dig.bench is a benchmark of 70 text-based games, 21 of them public, where an agent has to experiment its way to each game's unknown rules. Humans and frontier models play through identical game states, action sets and step budgets, and external testers confirmed every game is beatable by a human on a first attempt. The best models still struggle on the top tier.

CodeBurn is an open-source, local-first tracker that breaks down AI coding tokens and cost by task, model, tool and project across 40 tools including Claude Code, Cursor and Codex. It reads the session files those tools already write to disk, with no wrapper, proxy or API keys, and prices usage from LiteLLM rates refreshed daily. Its optimize command flags waste such as re-read files and unused MCP servers.