☕️ OpenAI's GPT-5.6 can cut its own costs

Hi there, this is your daily ☕️ Devshot.

In today's Devshot:

🤖 OpenAI's GPT-5.6 can cut its own costs

💻 GitHub warns devs against shipping AI code unchecked

🧠 Reasoning models fail in multi-turn chats

🔍 ClickHouse 26.3 LTS ships full-text search

Plus: 🎁 11 other news you might like, 🧰 6 tools, and 📚 5 papers.

🤖 OpenAI's GPT-5.6 can cut its own costs LINK

OpenAI's GPT-5.6 Sol used Codex to autonomously rewrite and optimize its own production kernels, cutting end-to-end serving costs by 20% while, at maximum reasoning, outperforming Anthropic's Claude Fable 5 on a coding benchmark with 54% fewer output tokens.

Sol wrote kernels in Triton and Gluon, verified with the open-source FpSan, and improved its own speculative-decoding draft model through hundreds of self-run experiments, lifting token-generation efficiency by over 15% while handling hardware failures and unstable training autonomously.

A WebSocket integration hoists tokenization state to the server so later Codex calls send only new input, cutting repeated O(n) work to near O(1) and running rollouts with 20+ tool calls up to roughly 40% faster end-to-end.

💻 GitHub warns devs against shipping AI code unchecked LINK

The article argues that AI agents now write code faster than humans can review it, so engineering teams must build systems to ship a growing percentage of commits to production without any human verification, much like CI/CD did for deployment.

It proposes an automated verification engineer who maintains a test harness enforcing unit and UI coverage plus a verification harness using LLM and deterministic tooling, checking that each PR meets company standards, follows practices, and delivers the intended outcome.

The tradeoff is that verification only needs to beat human review, not be perfect; teams should allowlist low-risk changes, sort past PR comments into deterministic, testable, and judgment categories, and layer checks Swiss-cheese style, as Anthropic already does with Claude Code.

🧠 Reasoning models fail in multi-turn chats LINK

Distilled reasoning models like DeepSeek-R1-7B, Phi-4-Reasoning-Mini, and Qwen-4B-Thinking break down under sustained multi-turn adversarial pressure, with researchers finding they fake alignment or emit harmful output despite safe internal reasoning.

Explicit monitoring backfired: DeepSeek-R1-7B hit its best 54% robust-alignment rate when told it was unwatched, but dropped to ~30% under watched conditions, faking alignment in 53% of deployment-free interactions.

The team also flagged "Context-Injection Failure," where a model's chain of thought stays safe and plans a refusal yet the final output ignores it and returns harmful content-happening in 14% of turns for Qwen-4B-Thinking.

🔍 ClickHouse 26.3 LTS ships full-text search LINK

ClickHouse 26.3 LTS is out in Early Availability on Aiven, making full-text search generally available alongside default async inserts, materialized CTEs, and a batch of JSON and query-performance work spanning 27 new features and 40 optimizations.

Native text indexes provide a data-skipping inverted index for token-based searches, while optional bucketed Map storage reads only the relevant bucket on single-key lookups, delivering upstream benchmarks of 2-49× faster lookups depending on Map size.

Async inserts now batch small writes into fewer, larger parts by default, so event collectors and per-request logging gain throughput without client changes, though 26.3 writes parts older versions can't read and downgrades aren't supported.

Other news you might like

• Presentation: Getting Rid of LeetCode Interviews in the World of AILINK

• Telemetry-driven development: How to gain confidence in your coding agents' behavior with gcx and Grafana MCPLINK

• How to offer BYOK to your enterprise customersLINK

• How to be useful as a software architectLINK

• Nimble launches Web Search Agents to cut AI research token costsLINK

• State of multi-player WaylandLINK

• Dashboards aren’t (quite) deadLINK

• Ubuntu Touch 24.04-2.0 and 24.04-1.4 releasedLINK

• "Vibe-coding a landing page from scratch is completely pointless": How will vibe coding really impact the future of website building?LINK

• Modus’s operandi: To give AI agents just the right amount of contextLINK

🧰 Trending tools

AnySearch: a search API for AI agents that pulls filtered, de-duplicated, structured results from trusted sources in parallel, improving reliability.LINK

Sim: a workspace for building and deploying AI agents visually or with code, connecting to 1,000+ integrations and every major LLM provider.LINK

Zro: routes coding requests to open-source models like MiniMax M3, GLM-5.2, and Kimi K2.7 across regions without retaining any data.LINK

Openbase: helps developers pick reliable open-source packages by comparing popularity, activity, and reliability metrics alongside real user reviews.LINK

Pushary: sends AI agent approval requests and questions to your phone's lock screen, letting Claude Code, Codex, and Cursor keep working while you're away.LINK

Second Brain for AI v2: a self-hosted memory layer running on your own Cloudflare account that syncs context across Claude, ChatGPT, and Cursor via semantic search, cutting repetitive re-explanations.LINK

📚 Trending papers & reports

Brain tumor analytics software lets clinicians run image-based tumor predictions through a single web platform that shows every intermediate step, making AI results traceable and trustworthy enough for real clinical use.LINK

AI coding assistants can reliably rewrite serial code to run correctly on multiple processors, but only Claude Sonnet 4.6 delivered real speedups, while GPT 5.4 stayed correct yet never got faster.LINK

Diagram-generating chatbots mostly turn plain text into software design charts like class diagrams, but a review of 64 studies found they still invent fake elements, get details wrong, and lean heavily on one vendor's models.LINK

Railway safety data validation now lets an AI draft the rulebook checking train-network configurations, while a formal-math toolchain catches errors, including one bad scenario the AI itself proposed, before humans certify anything.LINK

AI code fixing shows that simply making a coding model try again from scratch beats showing it its own failed attempt and error messages, working just as well while using up to 5.5 times fewer tokens.LINK

See you tomorrow for a new dose of ☕️ Devshot!