The Frontier Gets Cheaper Faster Than We Can Grade It
OpenAI, xAI and Meta all cut the price of intelligence this week, while two fresh audits show the cheap models we use to grade each other are shakier than the invoice suggests.
OpenAI splits GPT-5.6 into three standing tiers: Sol, Terra, Luna1
OpenAI began rolling out GPT-5.6 on July 9 not as one model but as three durable tiers: Sol the flagship, Terra a balanced default that matches GPT-5.5 at 2x lower cost, and Luna the cheapest.1 The ladder runs $1 in and $6 out per million tokens for Luna up to $5 and $30 for Sol, and the launch is a limited preview gated to trusted partners behind a US government safety review.2
Why it matters: the number now marks the generation and the name marks a capability tier that can advance on its own cadence. For anyone routing calls by cost, that means pinning a tier rather than a version, and re-pricing every time a tier ships.
xAI's Grok 4.5 undercuts Opus on the cost of writing code3
xAI shipped Grok 4.5 on July 8 as its first model built for coding and agentic work, at $2 input and $6 output per million tokens with a 500,000-token context. It sits third on the Artificial Analysis Intelligence Index across 168 models, and on SWE-Bench Pro it averaged 15,954 output tokens per task against 67,020 for Claude Opus 4.8, a launch-day figure that still needs production validation.3
Why it matters: token efficiency, not headline price, is where coding cost is now decided, and a fourfold gap in tokens per task swamps a per-token discount. The claim is xAI's own, so treat it as a hypothesis to benchmark against your codebase, not a settled result.
Meta moves its own AI accelerator into production to lean off Nvidia4
Meta's custom MTIA accelerator, code-named Iris, designed with Broadcom and built by TSMC, enters production in September; at least one chip cleared testing in about six weeks. Meta plans to deploy 7 gigawatts of compute this year and roughly double that in 2027, while still buying heavily from Nvidia and AMD.4
Why it matters: in-house silicon is a cost hedge against GPU shortages, aimed first at the ranking and recommendation models that pay Meta's bills. The GPU order book stays large, so this reads as diversification, not a break with Nvidia.
Swap the LLM judge and the score moves, with the answers unchanged5
A new audit finds an LLM-as-judge score shifts when you change the evaluator even though the candidate responses stay fixed, and reports that judge upgrades are not interchangeable across model families. Only scaling Qwen3 from 1.7B to 4B produced robust gains; stronger judges reduced but did not remove position and verbosity bias, and repeated-sample juries barely helped once errors were correlated.5
Why it matters: much of production evaluation now leans on a model grader. If a routine judge upgrade can move your leaderboard with no change to the system under test, your regression numbers are partly measuring the judge.
A frontier model is not required to verify citations6
Benchmarking eight judges across 1,248 rubric decisions, researchers found cheaper models hold up on deep-research source attribution: GPT-5-mini took the top source-relevance F1 at 0.908, and factual-support scores were statistically indistinguishable across the field.6
Why it matters: read against the judge-reliability audit, the lesson is not to reach for the biggest grader but to calibrate the one you pick. A cheap verifier you have audited beats an expensive one you trust on faith.
sqlite-utils 4.0 lands its first breaking release since 20207
Simon Willison shipped sqlite-utils 4.0 on July 7, its first major version
in nearly five years, adding schema migrations, nested transactions through
a new db.atomic() method, and compound foreign keys. He used
the window before Claude Fable left his subscription to drive the release
with a coding agent while holding to strict SemVer.7
Why it matters: a rare major bump built largely by an agent is a concrete marker of where agent-assisted maintenance is now dependable: mechanical, well-specified library work, with a human holding the version contract.
A one-line rule for spending less on coding agents8
Willison documented a project-memory instruction, credited to Jesse Vincent, telling Claude Code to route implementation into a subagent on a lower-power model while design, auditing and synthesis stay in the main loop. In practice that is Sonnet for substantive edits and Haiku for trivial ones, reserving the premium model for judgment.8
Why it matters: the cost lever most teams miss is not which model they pick but where they spend it. Delegating implementation down and keeping judgment up is the same routing logic the tiered price lists at the top of tonight's wire are built to reward.
Sources
- OpenAI Developer Community · Introducing GPT-5.6 series: Sol, Terra and Luna
- TestingCatalog · OpenAI launches GPT-5.6 Sol, Terra and Luna on apps and API
- WinBuzzer · Grok 4.5 targets coding workflows with lower API pricing
- TechCrunch · Meta's new AI chips will begin production in September
- arXiv · When the Judge Changes, So Does the Measurement: Auditing LLM-as-Judge Reliability
- arXiv · Do You Need a Frontier Model as a Citation Verifier? Benchmarking Rubric LLMs for Deep-Research Source Attribution
- Simon Willison · sqlite-utils 4.0, now with database schema migrations
- Simon Willison · Route coding implementation to a lower-power subagent