OpenAI's new GPT-6.2 costs you less

OpenAI's new GPT-6.2 models, Microsoft's MAI Code 2, and more.

OpenAI's new GPT-6.2 costs you less

Hi there, this is your daily ☕️ Devshot.

In today's Devshot:

🤖 OpenAI launches GPT-6.2 Astra and Sol

🗃️ Yelp moves Cassandra to EKS

📍 Turso Sync keeps SQLite writes local

🌳 Valkey 9.2 swaps skiplists for B+ trees

💻 Microsoft's MAI Code 2 rivals Claude

Plus: 🎁 7 more stories you might like, 🛠️ 5 tutorials, 🧰 6 tools, and 📚 5 papers.

OpenAI launches GPT-6.2 Astra and Sol LINK

  • OpenAI shipped two new models through ChatGPT and the API: GPT-6.2 Astra for demanding professional work, and GPT-6.2 Sol for faster everyday use, with both beating their predecessors on benchmarks at lower cost.
  • Astra API pricing drops 20% to $8 per million input tokens, while Sol holds at its predecessor's rate; Astra leads on science and legal benchmarks but coding scores flatlined and Sol slipped on code-quality tests.
  • Both models hit the "Critical" cybersecurity threshold in OpenAI's Preparedness Framework, prompting stricter safeguards that the company says will also limit legitimate vulnerability research, and Astra is the first released model to reach OpenAI's "High" bar for AI self-improvement.

Yelp moves Cassandra to EKS LINK

  • Yelp migrated roughly 90 production Cassandra clusters from self-managed Kubernetes to Amazon EKS with zero downtime and no data loss, re-pointing existing EBS volumes at new pods instead of streaming data between nodes over the network.
  • The "Rolling PV Transplant" exploits that EBS volumes bind to an Availability Zone, not a cluster: automation deletes one legacy pod at a time, waits for the volume to detach, and lets a pre-created EKS pod claim it and rejoin the ring.
  • About 1,000 TiB across ~1,000 pods moved by pointer rather than over the wire, with rollback just the reverse swap; Yelp warns teams to build explicit detach/attach timeouts and audit repair state, backup IAM roles, and alert severity first, after its own backups failed silently for 48 hours.

Turso Sync keeps SQLite writes local LINK

  • Turso Sync, the second attempt at SQLite replication from the team behind libSQL, now keeps writes on the local device by default and pushes them to Turso Cloud only when you explicitly call for it.
  • Built on Turso Database, a Rust rewrite of SQLite, it ships logical row-level changes via Change Data Capture instead of physical WAL pages, so a single-row insert costs about the size of that row and unrelated writes to different tables no longer collide.
  • Push and pull are separate calls and conflicts resolve as last-push-wins, which fits per-user data but breaks for counters or shared document fields where offline writes overwrite each other, so those cases need append-only designs or application merge logic.

Valkey 9.2 swaps skiplists for B+ trees LINK

  • Valkey 9.2 replaces the skiplist backing large sorted sets with a high-fanout B+ tree variant called fbtree, cutting per-member memory overhead by 36-43% in a multi-million-member test.
  • fbtree packs up to 61 entries per node versus the skiplist's effective branching factor of 4, stores each score and member together as one packed value, and uses SIMD feature comparisons plus linked leaves to cut dependent memory fetches.
  • Sorted set commands and behavior stay the same, though large sets now report fbtree rather than skiplist encoding, so update any tooling checking that literal string, and test the 9.2 release candidate against real large-sorted-set workloads before GA.

Microsoft's MAI Code 2 rivals Claude LINK

  • Microsoft unveiled MAI Code 2 at today's Ignite 2026 keynote, claiming the model matches 96% of Claude Opus 5.5's performance on everyday coding tasks while charging about 5% of its price via API.
  • Microsoft pitched MAI Code 2 as comparable to GPT-6.1 Sol or Opus 5.5, with a lighter Code 2 Flash variant rivaling GPT-6 Luna and Claude Haiku 5.5, though the reported pricing reflects a 50% introductory discount.
  • The models arrive as optional choices in GitHub Copilot, Microsoft 365, and Fabric, but Microsoft cautions they aren't built for the hardest agentic or research workloads, and the benchmark selection remains an open question.

Tutorials & Advice

> Deterministic Simulation Testing: Replace real clocks, networks, and disks with simulated ones running in a single loop, so every distributed-system failure becomes reproducible and fixable.

> Secret Weblog - How Rust Changed Me As A Developer: Learning Rust sharpens general programming skills by forcing clearer thinking about variant modeling, performance tradeoffs, and using types to guide design.

> Python Bindings: Calling C or C++ From Python: Explains how to call C or C++ functions from Python using binding tools so programs can reuse fast, stable libraries and bypass Python's speed limits.

> Understanding CSS random(): The new CSS random() function generates varied values like slightly off-kilter rotations for an organic look, defaulting to per-element scoping so results stay consistent across breakpoints.

> Is your OpenTelemetry pipeline becoming a performance bottleneck?: Reduce the resource cost of OpenTelemetry (the standard tooling for collecting app metrics, logs, and traces) by fixing batching, trimming unused attributes, and sampling traces.

Other news you might like

  • Running the full server-side Datasette Python app in the browser with service workers and Pyodide LINK
  • How Shopify Orchestrates AI-Powered Mobile App Development Without Overloading Developers' Machines LINK
  • Your AI coding agent could cost 5x more depending on the harness, benchmark finds LINK
  • RBRD: RAM-Backed Linux Block Driver Written In Rust For Better Safety But Slower Performance LINK
  • Flipper Devices launches $399 ‘Flipper One’ Linux-based multi-tool LINK
  • Ignite 2026: Microsoft Copilot SDK 2.0 is Now Available LINK
  • Building the Agent Host Protocol for sessions that continue across devices LINK

Trending tools

Jev: adds structured AI-driven decision-making to automation workflows, letting developers build reliable branching logic without hand-coding every rule LINK

Monospace from Directus: wraps any SQL database with instant REST and GraphQL APIs plus a no-code app for collaborative data management. LINK

iFixAi: audits AI agents with 250 inspections across 69 misalignment categories, flagging failures and giving engineers evidence to investigate problems LINK

Claude Opus 5.5: handles agentic coding and complex knowledge work with stronger alignment safeguards, running typical workloads at 40% less cost than Opus 5. LINK

Busabase: open-source shared workspace where teams and AI agents store records, docs, and apps with access controls and change tracking for reuse LINK

Claude Haiku 5.5: a fast, affordable small model from Anthropic for high-volume tasks like summarization, classification, coding subagents, customer support, and browser use. LINK

Trending papers & reports

Agent decision trails can be mined as evidence to flag the risky judgment calls AI coding agents make along the way, catching flawed assumptions that finished code and passing tests hide. LINK

AI coding-agent studies show that how researchers slice the data, not the code itself, often decides whether AI-written code outlasts human code, making many published verdicts unreliable. LINK

Virtual hardware testing lets coding agents build and debug embedded firmware without physical boards, revealing that even when code compiles ~99.7% of the time, ~20% of those firmwares still fail on realistic hardware checks. LINK

Coding-agent budgets work by hard limits, not conversation, since enforcing a token cap cut task success by 20 points while slashing spend twelvefold, yet telling the agent its budget never helped. LINK

Silent map errors can be caught automatically by checking whether an address converts back to its original coordinates, catching cases where Google confidently returned locations over 2km off for ~5% of a Puerto Rico aid delivery dataset. LINK

See you tomorrow for a new dose of ☕️ Devshot!

More from the archive