Model Selection

Vercel AI Gateway now exposes Qwen 3.8 Max to coding agents, adding one model endpoint for long-context text and vision work with gateway routing, budgets, and usage tracking.

Training-data curation can improve model quality and inference efficiency without simply adding compute. The practical work is decontamination, deduplication, balancing, task matching, and selective synthesis.

Base-model data is shifting from broad web imitation toward code, reasoning, and agent-task priors. The unresolved choice is how early to introduce synthetic and instruction-shaped data.

A multi-policy PEFT method groups and sequences compatible tasks into separate QLoRA paths, reducing interference without adding trainable capacity; it reached 44.78 on TRACE.

Coding benchmark scores can mislead when prompts leak implementation details or tests reject valid solutions. Inspect tasks and verifiers before using a leaderboard to choose an agent.

DeepSeek V4 Flash’s updated weights replace the preview behind the existing model ID, raising its reported Terminal-Bench score from 56.9 to 82.7 without code changes.

AI Gateway has raised Laguna S 2.1 capacity tenfold for both paid and free model IDs, reducing throughput constraints for high-volume or long-running coding agents.

Safety tuning against model self-consciousness may also shift unrelated value and mind-attribution responses. Builders using models for surveys or social reasoning should treat alignment as a confound.

OpenAI says GPT-5.6 Luna and Terra now cost less, which may change model-routing choices for agent workflows. The supplied material gives no prices or workload comparisons.

Vercel cut Luna and Terra token prices and raised Sol fast-mode speed without changing model IDs, so existing agent workloads inherit the changes without code edits.

MiniMax H3 brings short 2K video generation to Vercel AI Gateway, with text, keyframe, and multimodal reference inputs. Reference and keyframe modes cannot be combined.

Inkling Small is pitched as a lower-compute model for coding, tool use, and visual reasoning, with adjustable thinking effort and zero-data-retention routing through Vercel AI Gateway.

AI Gateway now exposes one beta fast-mode option across models, letting coding agents request lower latency while retaining standard-speed fallback when no fast tier exists.

OpenAI positions GPT-5.6 as delivering more useful output per dollar across inference and agent workflows. The supplied material has no metrics for judging routing or migration decisions.

ClinFusion combines native 2D and 3D medical-image understanding with region-grounded evaluation, offering a concrete architecture and eval design for clinical multimodal systems.

A 1,215-run study finds entity-matching architecture and model variant matter more than scale alone; generative matchers mainly help under distribution shift.

Kimi K3 now has US-hosted, ZDR-capable gateway routes plus a faster tier, giving coding-agent users explicit latency, residency, retention, and cost choices.

Edge-agent design is often memory-bound: use small prompted models when devices can spare gigabytes, or fine-tune tiny models for narrow functions when reach and latency matter more.

The same model identifier produced sharply different judgments across API and web deployments. Treat model, interface, system configuration, and date as one versioned dependency.

TRACE-Router selects one model per agent task, keeps every call on that backend, and learns from the final outcome. Its benchmarks suggest task-level routing can improve accuracy and latency together.

AI Gateway now serves Claude Opus 5 with configurable reasoning, fast mode, fallbacks, and coding-agent setup; benign security tasks may still hit safeguards.

Agent economics can regress even when token prices look stable. Route by task, preserve model optionality, and move deterministic work out of LLM calls before scaling usage.

Ling 3.0 Flash joins AI Gateway with a 256K context window, thinking and non-thinking modes, and free access through August 3 for agent workload testing.

Cursor Router classifies coding requests and selects models by task and cost. Cursor reports lower spend in production tests, but the strongest evidence is limited to its own traffic and metrics.

Poolside’s open-weight Laguna S 2.1 is now callable through Vercel AI Gateway in free 256K and paid 1M-context variants, with thinking and no-thinking modes for coding agents.

Vercel AI Gateway added Gemini 3.6 Flash for coding and web work and Gemini 3.5 Flash-Lite for scoped subagent tasks, both available through AI SDK model IDs.

Vercel’s June gateway data shows cheap volume moving to open-weight models while costly agent workloads stay on frontier models, strengthening the case for risk-based routing.

Claude Fable 5 is back on Vercel's AI Gateway after US export controls lifted. New, stricter safety classifiers can refuse routine coding calls, so configure model fallbacks; no Zero Data Retention (30-day hold).

Cursor audited SWE-bench runs: 63% of Opus 4.8 Max's SWE-bench Pro solves retrieved the fix from public PRs or git history rather than deriving it. Sealed harnesses cut scores by up to 20 points.

Vercel AI Gateway adds firewall-style routing rules: rewrite one model to another or deny a model outright, applied at the gateway so you swap models across your whole team without shipping a code change.

Anthropic details Fable 5's cyber classifiers — pen testing and exploit dev are blocked even for legitimate use, with a wider false-positive margin — and drafts a five-level jailbreak severity scale (CJS).

Sonnet 5 lands June 30 at $2/$10 per Mtok intro pricing (through Aug 31), nearing Opus 4.8 on agentic coding and computer use. Note the new tokenizer: inputs map to 1.0–1.35x more tokens.

Fable 5 is back globally as of July 1 after US export controls (June 12–30) triggered by an Amazon-discovered jailbreak; a new classifier blocks the technique in over 99% of cases.

The US government ordered Anthropic to suspend Fable 5 and Mythos 5 globally, citing a jailbreak claim Anthropic disputes. If your agent stack pins either model, you need a fallback now.

Program-as-Weights compiles natural-language fuzzy functions (JSON repair, log filtering) into adapters for a frozen 0.6B interpreter — matching Qwen3-32B prompting at ~1/50th the memory, 30 tok/s on an M3.

A Dharma AI essay on Goldfeder, Wyder, LeCun and Shwartz-Ziv (2026) argues specialized models beat generalists under fixed resources — a case for narrow models and task-scoped agents over one generalist.

Every Eval Ever's ~229k benchmark results across 22k+ models now cross-post to Hugging Face model pages with attribution — one less reason for the same model to show two different MMLU scores.

Treble and Hugging Face launched FFASR, a leaderboard testing ASR models in simulated far-field rooms at three SNR bands. If you build voice interfaces for agents, near-field WER numbers oversell real-room accuracy.

Cursor is convening a CFO Council on AI ROI, and its telemetry is the useful part: cost per agent request varies ~9x across model families, and 84% of power users run multiple models weekly.

Vercel’s AI Gateway rankings are now queryable as CSV or JSON, giving agent builders production-usage signals for model and provider selection beyond benchmark scores.