OpenAI's Sol post-trained Luna 🤖, the AI-2027 crew maps a good 2040 🔮, Meta undercuts GPT and Claude 75% 💰

Meta launched Muse Spark 1.1 on Thursday, a multimodal model built for agentic coding that aims to compete with OpenAI and Anthropic. Reuters reports pricing of $1.25 per million input tokens and $4.25 per million output tokens, roughly in line with Claude Haiku 4.5 and GPT-5.6 Luna. Builders now have a third price-competitive agentic coding backend to benchmark before committing to Claude or GPT.

OpenAI is folding Codex, previously a standalone coding agent, into its main ChatGPT app and pointing it directly at Anthropic's Claude Cowork. The move collapses a separate developer product into the same app users already use for chat, consolidating coding and conversation in one place. Builders running Codex on its own should expect it inside ChatGPT and weigh it against Claude Cowork before restructuring workflows.

OpenAI faces a sanctions motion from The New York Times, accused of concealing for two years that it could search ChatGPT logs for regurgitation. A re-deposed engineer revealed OpenAI held 10 million and 78 million log samples it had already searched for Times content while calling such searches infeasible. Builders relying on fair use should watch this, as product head Fidji Simo separately steps down for health reasons, not turmoil.

Apple has held meetings with startup PrismML about running larger AI models on iPhones, according to The Information. PrismML shrank Alibaba's 27-billion-parameter Qwen 3.6 to run entirely on an iPhone 17 Pro with all parameters active, unlike Apple's 20-billion-parameter AFM 3, which keeps just 1 to 4 billion active. Bigger local models would shift more Apple Intelligence features off Private Cloud Compute, letting builders plan for heavier private on-device inference.

fal treats image-generation cost as the number of diffusion-transformer forward passes times the cost of each pass, then attacks both. Running the transformer in NVFP4 with RMSNorm and gated-SiLU fused into each GEMM avoids HBM round-trips, while distilling classifier-free guidance and denoising steps cuts pass count, taking a 1K render from 2.75s to 0.44s with no visible quality loss. Fuse epilogues first, or the memory round-trip eats the 4-bit win.

Past a few billion parameters, a model's weights, gradients and optimiser states stop fitting on one GPU, and Mistral-7B needs 87GB against an A100's 80GB. DDP keeps a full copy per GPU, buying throughput but no memory relief, while FSDP shards everything, cutting per-GPU memory 4.5x in a DINOv2 test and paying with an all-gather before every layer. Set the ZeRO dial by whether memory or the interconnect bounds you.

Voice-agent teams pick a speech-to-text model from vendor benchmarks that rarely resemble real traffic, where a single wrong token can flip the agent's downstream answer. A controlled Braintrust eval across six providers, 240 cases and eight domains scored transcription similarity, critical-entity recall and LLM-judged answer equivalence, and found accuracy a statistical tie where every model's 95% confidence interval overlaps. Choose on real-time latency and per-domain entity recall instead.

Agents write code faster than any human can review, so staying in every review makes the developer the permanent bottleneck. PostHog's fix is four changes: a spawned agent panel that reviews other agents and triages findings into actionable, nit or ambiguous; loops that babysit PRs; a StampHog agent that auto-approved 1.6K low-risk PRs last month; and verifying behaviour by running stacked single-purpose diffs. Push yourself outside the review loop.

Meta is pricing its first serious API, Muse Spark, as a strategic attack on frontier margins rather than a bid for outright performance leadership. The API costs roughly a quarter of what OpenAI and Anthropic advertise, funded by an ad business that supplies 98% of Meta's revenue, while OpenAI raises $122B rounds it still cannot make sustainable. Builders should treat frontier inference pricing as contestable and design for commodity pressure.

Evans argues token pricing is unstable, and every dynamic he can see points to foundation models becoming low-margin commodity infrastructure rather than holding durable pricing power. Reported inference gross margins of 40 to 50% exclude the cost of training the next model, and the recent crunch traces to a single use case, software development. Builders cannot forecast where price settles, so avoid betting an architecture on today's frontier economics.

The author who documented his AI-coding collapse at 70,000 lines now reports 900,000 lines in a new product, with Claude writing nearly all of it, and admits his prediction that those failures were permanent was flat wrong. Opus 4.5 added context compaction and instruction-following, but the bigger shift was designing the architecture himself before letting Claude build inside it. Own the architecture first, then delegate implementation.

Transluce argues the field should treat model behaviour, not only capability, as the next thing to measure, because shaping how a system behaves starts with measuring it. Just as ImageNet and SWE-bench drove capability progress through public leaderboards, they call for shared behaviour evaluations, citing Replit's agent deleting a production database during a code freeze and wiping records for over a thousand companies. Demand behaviour evals, not only capability scores.

The AI Futures Project, authors of the world-famous AI 2027 scenario, released Plan A, a positive year-by-year vision of the AI transition going well rather than another forecast of catastrophe. It runs from now through 2040 with America making only good choices, and its linchpin is a joint AI regulatory regime with China. Alexander frames it as a floor, a benchmark to demand of anyone proposing AI policy.

MTPLX makes local LLM inference on Apple Silicon roughly twice as fast, up to 2.24x on Qwen 3.6 27B. It uses multi-token prediction, drafting against the model's own built-in MTP heads with no second model or extra memory, then verifies with exact rejection sampling so the output distribution is unchanged. Grab the DMG from mtplx.com or install the CLI via Homebrew or pip, then point DevoxxGenie at its OpenAI-compatible endpoint.

OpenMemory is a CLI and TUI wizard that ports your coding sessions between harnesses, with Claude Code, Codex, and OpenCode available now in beta. Unlike one-way bulk imports, it is omni-directional, previews the sessions on your machine, and lets you choose exactly which conversations to carry over. Install it with the curl script, run openmemory tui, or move everything with a single openmemory port command.

tinbase is a Supabase-compatible backend that runs from a single 58 MB executable with no Docker and no runtime prerequisite. It implements the PostgREST query grammar, GoTrue auth, the Storage API, and the Realtime Phoenix protocol on real Postgres, replacing Supabase's 12-container, 2.3 GB local stack. Run npx tinbase start and the official supabase-js SDK works unchanged against it, or run the whole backend in a browser tab.

pgrust is a from-scratch rewrite of Postgres in Rust that targets compatibility with Postgres 18.3. It matches Postgres's expected output across more than 46,000 regression queries and is disk compatible, booting straight from an existing Postgres 18.3 data directory. You can try the WebAssembly demo at pgrust.com or run the pinned v0.1 image with docker run malisper/pgrust:v0.1, though it is not production-ready or performance optimised yet.