AI Model Roundup: June 30–July 6, 2026

This week saw three major launches—Anthropic's Claude Sonnet 5, OpenAI's gated preview of GPT-5.6 Sol/Terra/Luna, and Meituan's open-source release of the massive LongCat 2.0 coding model. We also got new diffusion models from Google and NVIDIA, a pricing shakeup from DeepSeek, and Anthropic quietly redeployed Claude Fable 5 after last month's regulatory shutdown. Here's what mattered.

Anthropic Launches Claude Sonnet 5—With a Cost Catch

On June 30, Anthropic released Claude Sonnet 5, positioning it as its most agentic midrange model. It's now the default across all Claude plans and available via API as claude-sonnet-5. Anthropic's blog claims:

• SWE-Bench Pro: 5.1% improvement over Sonnet 4.6
• Terminal-Bench 2.1: 13.4% improvement over Sonnet 4.6
• GDPval-AA v2: scored 1,618, up from Sonnet 4.5's 1,395

Pricing: Introductory $2/M input, $10/M output through August 31; then $3/M input, $15/M output. But that's not the full story. Sonnet 5 uses a new tokenizer that inflates input tokens by 1.0–1.35x depending on content. Combined with more aggressive agentic behavior (~40% more output tokens per task at max settings), independent analysis from Artificial Analysis found Sonnet 5 costs $2.29 per task on average—more than Opus 4.6's ~$1.97. The Decoder called it a pattern of hiding price increases behind unchanged token rates.

Separately, Anthropic also redeployed Claude Fable 5 after US export restrictions were lifted. BBC reported the ban, imposed in early June under new AI export rules, was rescinded following Anthropic's compliance with updated security requirements. Fable 5 is now available again to enterprise customers in allied nations.

OpenAI Previews GPT-5.6 Sol, Terra, and Luna

OpenAI released a limited preview of GPT-5.6 in three variants: Sol (flagship), Terra (high-efficiency), and Luna (lightweight). OpenAI's announcement frames this as a government-gated release—only approved partners can access them, per US regulatory requirements. VentureBeat confirmed the access restrictions.

The GPT-5.6 Preview System Card provides technical details: Sol is a dense architecture with chain-of-thought reasoning, while Terra uses a MoE variant optimized for throughput. Benchmark numbers are sparse—OpenAI hasn't released standard evaluation scores—but the help center article notes Sol outperforms GPT-5 on internal coding and safety evaluations.

No pricing or general availability dates were provided. This is a controlled rollout, likely aimed at government and defense contractors first.

Notably, SiliconANGLE reports GPT-5.6 Terra outperforms Sonnet 5 on Terminal-Bench 2.1 by ~4%—one of the few cross-model comparisons available this week.

Meituan Open-Sources LongCat 2.0: 1.6T Parameters, No NVIDIA GPUs

Meituan open-sourced LongCat 2.0, a 1.6-trillion-parameter agentic coding model trained entirely on Chinese chips. VentureBeat reports it has been leading agentic coding leaderboards on OpenRouter, outperforming models like DeepSeek V4 Pro and Qwen3.7 Max on coding benchmarks. The model is available on Hugging Face under a permissive license.

This is significant for two reasons: (1) it demonstrates that frontier-level coding performance is achievable without US hardware, and (2) at 1.6T parameters, it's one of the largest open-weight models ever released. Expect this to accelerate the open-weight coding arms race.

For context, we covered the initial LongCat 2.0 story earlier—see our previous post: LongCat-2.0: China's 1.6T-Parameter Coding Model Trained Without a Single Nvidia GPU.

New Diffusion Models: DiffusionGemma and Nemotron Labs Diffusion

Two notable diffusion model releases this week:

• DiffusionGemma from Google DeepMind: A 26B-parameter (4B active) text-to-image model using a diffusion architecture rather than autoregressive. Google's blog claims 4x faster generation than equivalently sized autoregressive models. Available on Hugging Face: google/diffusiongemma-26B-A4B-it. The model overview notes it's optimized for on-device deployment.
• Nemotron Labs Diffusion from NVIDIA: An open-source diffusion model for text-to-image generation. GitHub repository includes trained weights and training code. No benchmark numbers published yet.

Pricing Updates and Benchmark Shifts

DeepSeek pricing change: As of July 1, DeepSeek updated its API pricing. Current rates show DeepSeek V4 Pro at $0.50/M input, $2.00/M output—a 20% reduction from previous pricing. The move continues the price war trend we covered in The AI Price War of 2026.

ByteDance EdgeBench: ByteDance Seed released EdgeBench, a benchmark for evaluating small models on edge devices. The accompanying dataset on Hugging Face covers latency, memory, and accuracy tradeoffs across 15 edge-relevant tasks. Early results show Seed-2.0-Lite leading in accuracy-per-watt among sub-7B models.

Model	Input Price ($/MTok)	Output Price ($/MTok)	Key Benchmark Gain
Claude Sonnet 5	$2 (intro), then $3	$10 (intro), then $15	+5.1% SWE-Bench Pro vs Sonnet 4.6
DeepSeek V4 Pro	$0.50 (new)	$2.00 (new)	20% price reduction
LongCat 2.0	Open-weight (free)	Open-weight (free)	Leading OpenRouter agentic coding

Key Links and Source References

• Anthropic: Introducing Claude Sonnet 5
• OpenAI: Previewing GPT-5.6 Sol
• OpenAI Help Center: GPT-5.6 Sol, Terra, and Luna preview
• GPT-5.6 System Card
• Anthropic: Redeploying Claude Fable 5
• BBC: US lifts export ban on Anthropic's advanced AI tools
• VentureBeat: Meituan open-sources LongCat 2.0
• Google: Introducing DiffusionGemma
• NVIDIA: Nemotron-Labs-Diffusion on GitHub
• DeepSeek API Pricing
• ByteDance EdgeBench GitHub

Bottom line: This week's biggest news is the hidden cost of Sonnet 5—lower token prices don't matter if the model uses 40% more tokens per task. Meanwhile, LongCat 2.0's open-source release and DeepSeek's price cuts put pressure on proprietary pricing. The GPT-5.6 preview remains behind closed doors; watch for benchmark leaks in the weeks ahead.