AI DIGEST: 29 JUNE – 5 JULY 2026

Five things from the turn of the month, picked out of everything we published on the daily feed.

A Crowded Week of Model Releases

OpenAI opened a limited preview of GPT-5.6 in three sizes (Sol, Terra and Luna), with Luna priced from $1 per million input tokens and a system card emphasising cyber and bio testing. Anthropic released Claude Sonnet 5, a cheaper model approaching Opus 4.8 on planning, tool use and coding, and restored Fable 5 and Mythos 5 access after US export controls were lifted. Apple shipped five third-generation foundation models split across on-device and Private Cloud Compute. The pattern worth tracking is not the flagships but the tiers beneath them: the mid-range is where the price-performance of production workloads is actually being decided.

Long-Horizon Software Work Is Still Unsolved

RoadmapBench evaluates coding agents on real version-upgrade tasks (a median of 3,700 changed lines across 51 files), and 13 frontier models scored between 5.2% and 39.1%. Epoch AI and METR launched MirrorCode, which asks whether a model can reimplement an entire real program end to end without source access, the leader managed 56%. Set against the daily experience of agents closing tickets competently, these numbers are the useful corrective: the difficulty is not writing code, it is holding a large, evolving change together over many steps. Scope your agent work to bounded, well-specified changes, and keep a human owning the shape of anything that spans a repository.

Routing Beat Scaling on Cost

Cognition's Devin Fusion combines frontier and cheaper models in a dual-agent design that routes each task dynamically, cutting cost 35% on the FrontierCode benchmark with no loss of top-tier performance, and 41% once Fable 5 was in the mix. The same instinct, spending the expensive resource only where it pays, shows up lower in the stack, with Moondream's write-up on closing GPU bubbles by overlapping CPU and GPU work between tokens, and a careful benchmark of KV cache quantisation showing that perplexity hides the tail damage which actually breaks tool calls and structured output. Routing is the highest-leverage cost lever most teams have not pulled, and it needs no model changes at all.

Custom Models Beat the Frontier on Proprietary Work

Thinking Machines reported that frontier models struggle with relatively simple financial tasks, while smaller custom models trained on proprietary data labelled by expert investors do better at lower cost. A companion argument makes the general case that finite capacity favours concentration over universality, and poolside's Laguna XS 2.1, a 33B MoE aimed squarely at agentic coding, is the same bet in product form. The practical implication is about data, not modelling: the asset that makes a specialised model possible is a well-labelled record of expert judgement, and most organisations are throwing theirs away daily.

A Big Context Window Is Not Memory

A clear explainer this week set out why 1M-token contexts do not give agents memory: agents still repeat themselves, lose the thread of long-running tasks and forget decisions, because holding text is not the same as knowing what matters. A research paper made the constructive version of the point, treating memory as reconstruction rather than retrieval: an associative graph of cues and content that the agent explores and prunes as evidence accumulates, improving accuracy on long-history benchmarks while reducing tokens and runtime. If you are solving a memory problem by buying a larger context window, you are paying more for the same failure.