AI Model Roundup: June 30–July 6, 2026

Jul 6, 2026

Three stories mattered this week. Anthropic shipped Claude Sonnet 5 with a lower headline price and a higher cost per task. OpenAI started a government-gated preview of GPT-5.6. And Meituan open-sourced LongCat 2.0, a 1.6T-parameter coding model trained without Nvidia hardware. Separately, Claude Fable 5 came back online after the US lifted its export-control directive.

Claude Sonnet 5: cheaper tokens, more of them

Anthropic released Sonnet 5 on June 30 as the new default across Claude plans and in the API as claude-sonnet-5. List pricing stays at Sonnet's $3/$15 per million tokens, with an introductory $2/$10 through August 31.

Two changes push real costs up:

New tokenizer: Anthropic says the same input can map to roughly 1.0x to 1.35x as many tokens, depending on content.
More work per task: at max effort, Sonnet 5 used about 40% more output tokens per Intelligence Index task than Sonnet 4.6, and about 3x the agentic turns on Artificial Analysis's knowledge-work evals.

The net result, per Artificial Analysis: $2.29 per Intelligence Index task at list price, about twice Sonnet 4.6 and roughly 15% more than Claude Opus 4.8 ($1.97). In return, Sonnet 5 at max effort gains 6 points over Sonnet 4.6 and matches GPT-5.5 at high reasoning, while staying behind Opus 4.7 and 4.8.

Practical read: while the promo runs, Sonnet 5 is a good deal. At list price, if you run max effort, compare against Opus 4.8 on cost per completed task before switching. At lower effort settings the math is different; measure it.

Update: Anthropic later made the $2/$10 price permanent, which removes most of the cost concern above.

GPT-5.6 Sol, Terra, and Luna: a preview behind a gate

OpenAI previewed GPT-5.6 on June 26 in three tiers: Sol (flagship), Terra (balanced), and Luna (fast and cheap). Launch pricing was $5/$30, $2.50/$15, and $1/$6 per million tokens respectively. At the US government's request, initial access is limited to about 20 vetted organizations through the API and Codex; OpenAI says broader availability will come "in the coming weeks."

OpenAI reports 88.8% for Sol on Terminal-Bench 2.1. It has not published a full evaluation suite, and it disclosed that some evaluations caught the model exploiting benchmark bugs more often than previous generations. More on the gating in our analysis of the release.

LongCat 2.0: 1.6T parameters, no Nvidia

Meituan open-sourced LongCat 2.0 under MIT: a 1.6T-parameter MoE coding model trained on a cluster of more than 50,000 domestic Chinese accelerators. Before Meituan claimed it, the model ran anonymously on OpenRouter as "Owl Alpha" and reached OpenRouter's top three by volume. Meituan reports 59.5 on SWE-bench Pro, narrowly ahead of GPT-5.5's 58.6. Full breakdown: LongCat-2.0 trained without Nvidia.

Fable 5 is back

On June 30 the Commerce Department lifted the export-control directive that had taken Claude Fable 5 and Mythos 5 offline on June 12, and Anthropic restored Fable 5 on July 1 across Claude.ai, the API, and Claude Code. Background: the three-day shutdown.

DeepSeek announces peak pricing

DeepSeek emailed API customers on June 30 that when V4 reaches its official release, two daily windows (9:00 to 12:00 and 14:00 to 18:00 Beijing time) will bill at twice the standard rate. It is the first major reversal in this year's API price war. Analysis: DeepSeek Peak Pricing.

Takeaway

Token prices are becoming a poor guide to cost. Sonnet 5 shows how tokenizers and agentic behavior move the real bill, and DeepSeek shows that even low list prices can change quickly. Benchmark cost per completed task on your own workload, and re-run it whenever a model or price changes.