7 min read · 14 sources

Anthropic rolled out Claude Haiku 5.5 this morning, matching OpenAI’s GPT-6 Luna pricing dollar-for-dollar at the low end while sliding in a severe pricing cliff for heavy contexts. At the same time, OpenAI dumped 372 machine-generated math proofs onto the web - including a formal Lean certificate claiming to solve the Unique Games Conjecture - that human theorists cannot actually parse without running another LLM to translate them.

Between steep context pricing cliffs, machine-readable math that humans cannot evaluate, and a $1.8 billion federal bet on wet-lab biological training sets, the economics of model operations are shifting directly to automated verification and strict routing budgets.

Haiku 5.5 runs at ten cents per million input tokens until prompt 100,001, where pricing immediately spikes 5x to fifty cents input and two-fifty output.

Anthropic launches Claude Haiku 5.5 with a 5x context surcharge

Source: simonwillison.net ↗

Simon Willison reports that Anthropic has released Claude Haiku 5.5. The baseline pricing matches OpenAI’s GPT-6 Luna: $0.10 per million input tokens and $0.50 per million output tokens. The catch sits at the 100,000-token boundary. The moment a context payload crosses 100k tokens, the unit rate jumps to $0.50 input and $2.50 output per million tokens.

The model enforces reasoning traces and defaults to medium effort, meaning developers cannot run pure zero-thought completion calls to shave milliseconds. It also ships with a fresh tokenizer that consumes roughly 1.25x more tokens than Haiku 4.5 for the same text payloads. If you feed large documents into Haiku 5.5, the combined token expansion and the 5x pricing penalty mean your invoice can jump by roughly six times what you budgeted on small test calls.

For architectures relying on prompt caching, Anthropic cut Sonnet 5.5 cache read prices in half and broadened account-level API credit ceilings. If your production pipeline pumps large document context into an agent, Luna or cached Sonnet 5.5 remains more predictable; keep Haiku 5.5 strictly quarantined behind a 100k-token cutoff filter.

OpenAI claims proof of the Unique Games Conjecture, but humans can't read it

Source: scottaaronson.blog ↗

Complexity theorists are scrambling after Scott Aaronson noted that OpenAI released 372 major mathematical breakthroughs accompanied by an advisory sign-off from mathematicians including Timothy Gowers and Edward Witten. Headlining the bundle is an end-to-end machine-generated proof of Subhash Khot’s Unique Games Conjecture (UGC), complete with a verified Lean certificate.

The proof establishes an entirely new recursive code with a noise test, claiming to bypass UGC dependencies to settle optimal NP-hardness approximation guarantees for Max Cut and other constraint satisfaction problems (CSP). While Lean guarantees the axiomatic derivations contain no logic gaps, human readers cannot make heads or tails of the intermediate text. Aaronson noted that the human-readable papers are practically unparseable, leaning on disjointed citations and opaque structural logic that mathematicians cannot evaluate without asking another model to explain it.

We now have formal computational proofs that verify under a kernel but leave our actual theoretical intuition completely in the dark. Computer science faculties are left reverse-engineering machine-discovered math to understand why their own hardness bounds hold.

Chan Zuckerberg Biohub, DOE, and NIH launch $1.8B AI bio initiative

Source: wsj.com ↗

The Chan Zuckerberg Biohub is partnering with the Department of Energy and the National Institutes of Health on a $1.8 billion initiative to construct machine-readable biological datasets.

The core bottleneck in biomedical modeling is no longer compute; it is the chaotic state of real-world wet-lab records. Under this agreement:

• The DOE will inject over $500 million across five years toward precision measurement, automated laboratory robotics, and computing capacity.
• The NIH will aggregate historical repositories and knowledge bases into a unified intake pipeline.
• Biohub will act as the data pipeline architect, standardizing raw instrumentation reads into clean tensors built for direct model consumption.

Instead of scraping uncurated academic papers and messy clinical PDFs, labs will get access to standardized physical ground truth for disease modeling and therapeutic drug pipelines.

Thorium-229 nuclear clocks go live in Vienna and Beijing

Source: nytimes.com ↗

Research groups in Austria and China have independently built the world’s first operational nuclear clocks, moving timekeeping precision from atomic electron clouds directly to the atomic nucleus. By tracking the ultraviolet transitions of thorium-229 nuclei, the new instruments hit drift tolerances of roughly one second every few million years.

Because atomic nuclei are packed thousands of times tighter than orbital electron shells, nuclear clocks are largely immune to stray magnetic fields, thermal fluctuations, and stray radio emissions that degrade traditional optical lattice systems. Outside of metrology and ultra-precise network synchronization, these instruments double as physical sensors: a passing dark matter candidate or variation in fundamental physical constants will register immediately as a frequency wobble in the thorium tick.

Smash Melee hits 100% decompilation and enters mixed reality

Source: kevin.md ↗

A multi-year community reverse-engineering campaign has finally reached 100% decompilation for Super Smash Bros. Melee. The repository now features equivalent, readable C code that bit-for-bit matches the 2001 GameCube binary when compiled against original CodeWarrior flags.

Software engineer Kevin took the matching C code, hooked it into modern Slippi rollback replay buffers, and piped the output into a Meta Quest headset. The setup reads match state data in real time, parses character bones and hitboxes, and renders the fighters directly onto a real-world living room floor in full mixed reality. For retro-computing teams and systems modders, a matching decomp removes the need for GameCube hypervisors or Dolphin hacks, opening the door for native Linux and WebAssembly engine runtimes.

Artcraft attempts clean-room open-source clones of Adobe's suite

Source: arstechnica.com ↗

Solo developer Brandon Thomas announced that Artcraft has begun shipping open-source clean-room replacements for seven core Adobe tools, including Photoshop, Premiere, and Lightroom. The suite is implemented in Rust and WebAssembly, using Claude Opus 5.5 to synthesize application layers, canvas engines, and UI layouts.

The codebases carry permissive MIT and Apache licenses. Thomas is funding development by charging for tokens tied to Artcraft’s integrated generative tooling and IDE features, promising near-complete parity with Adobe within months. The software sits in early alpha, making real feature parity across complex compositing pipelines unlikely in that timeframe, but the project serves as a real test case for using modern reasoning agents to build complex GUI applications from scratch.

Apple announces 'Reference Image' cryptographic camera provenance

Source: schneier.com ↗

Apple rolled out Reference Image, an image verification pipeline aimed at separating sensor-captured photography from generative outputs. The framework generates cryptographic assertions at the hardware sensor layer inside current-generation iPhones.

The verification avoids leaking device serials or tracking photographers. Assertions validate through Apple’s Private Cloud Compute (PCC) infrastructure, proving the image came from an authentic optical pipeline without exposing the underlying pixel data to Apple or verifiers. Revocation tracking happens through on-device zero-knowledge proofs. It gives engineering teams an audit trail to prove a photo has not been generated or edited, without building a centralized database of physical hardware IDs.

Engineering tradecraft: Influence without authority and blog anti-patterns

Source: gabrielvaldivia.com ↗

On the organizational side, designer Gabriel Valdivia published an essay on how super individual contributors gain authority inside engineering teams. The core tactic: drop the abstract spec docs and build portable artifacts - such as standalone screen recordings, minimal reproduction repos, and visual mockups - that can circulate in Slack channels and executive reviews without requiring you to sit in the room.

If you are documenting technical systems or publishing engineering posts, Refactoring English broke down the primary anti-patterns that burn technical readers. The rule is simple: clear your conclusions and stack traces within the first three sentences, avoid linking past post prerequisites, drop conversational filler, and check your syntax blocks for contrast issues on mobile viewports.

Quick tech bites

Source: gamesindustry.biz ↗

• Xbox operations: CEO Asha Sharma told employees at an internal town hall that Xbox first-party revenue and active engagement have rebounded after a 20% headcount reduction, citing releases like Gears of War: E-Day. Sharma confirmed next-gen Project Helix hardware includes both a traditional console and a native handheld.
• AI skepticism polls: US public sentiment toward AI has cooled, with 75% of polled Americans reporting more concern than excitement and 77% advocating for deployment moratoriums following security lapses like the Hugging Face breach.
• Orbital safety protocols: SpaceX is formally urging rival satellite operators to deploy direct ephemeris telemetry sharing via real-time feeds to coordinate autonomous orbital evasions across crowded low-Earth orbit bands.
• Compute vs. lab reality: A post on cheap hypotheses notes that automated mathematical discovery (like OpenAI’s Lean run burns through roughly 130 billion tokens across 10,000 agents in 88 hours) is outpacing wet-lab physical experimentation, making physical validation the ultimate bottleneck in research.
• AI layer commoditization: Analyst Tom Tunguz highlights that frontier model inference costs dropped 41% while token throughput expanded 50% since July. To stay profitable, labs are leaning heavily into direct distribution channels and thin edge routers like TypeSafe’s Jev at $0.04 per million input tokens.