Model Release Changelog — July 2026

Recurring format: Monthly roundup of frontier closed-model releases, open-weight drops, pricing changes, and practical recommendations. June 2026 edition · You are here: July 2026

July 2026 at a glance

Frontier closed models

Anthropic — Claude Opus 4.7 GA (July 2)

June's biggest tease delivered.

What shipped:

• Claude Opus 4.7 — general API availability (was limited-access preview in late June).
• Claude Code adds /model opus as a first-class switch for plan mode and hard refactors.
• Extended thinking budget increased 20% on Opus without extra latency on standard queries.

Benchmark highlights:

• GPQA Diamond: 84.2% (up from 77.1% on Opus 4.5).
• Multi-step agent plans (internal Anthropic eval): +9% completion rate vs Sonnet 4.7.
• SWE-bench Verified: 74.8% — marginal gain over Sonnet 4.7 (72.1%); not the reason to switch.

Pricing: Unchanged. Opus $15/$75 per 1M tokens. Sonnet $3/$15.

What to do: Add Opus to your routing layer for the hardest 10% of queries — architecture reviews, gnarly debugging, eval design. Keep Sonnet as default for Claude Code and Cursor daily loops. See Best AI Model for Coding.

OpenAI — GPT-5 Pro price cut (July 8)

Pro tier joins the price war.

What shipped:

• GPT-5 Pro — 25% price reduction across API and ChatGPT Team API access.
• Batch API discount extended to GPT-5 Pro (additional 50% off batch jobs).
• Responses API: native support for tool-use + structured output in one call (all GPT-5 variants).

Benchmark highlights:

• No new capability claims — this is a pricing release.
• GPT-5 Pro MMLU: 91.4% (unchanged from June).
• Reasoning mode on Pro: still 2x slower than GPT-5.1 standard mode — use deliberately.

Pricing: GPT-5 Pro $7.50/$30 per 1M tokens (was $10/$40). GPT-5.1 and Mini unchanged from June.

What to do: Re-run cost comparisons for workloads you kept on Sonnet because Pro was too expensive — legal doc analysis, long report generation, and multi-step planning may now fit Pro budget. Update LLM cost optimization routing tables.

Google — Gemini 2.5 Ultra (July 15)

Google's multimodal agent play.

What shipped:

• Gemini 2.5 Ultra — native video, audio, and screenshot understanding in one model.
• Agent mode in AI Studio — tool-use loop with browser and code execution sandbox.
• Vertex AI GA for Ultra on enterprise contracts; consumer access via Gemini Advanced.

Benchmark highlights:

• Video QA (internal Google eval): 89% accuracy on 5-minute screen recordings.
• Tool-use reliability: +14% over Gemini 2.5 Pro on multi-step web tasks.
• Context: 1M tokens (Ultra trades max context vs Pro's 2M for modality breadth).

Pricing: Ultra $3.50/$14 per 1M tokens — premium over Pro ($1.25/$5) but under Opus.

What to do: Use Ultra for multimodal agent prototypes — UI testing from screenshots, meeting recording summarization, video tutorial Q&A. Keep Pro for text-only long-doc workflows. See GPT vs Claude vs Gemini.

Cohere — Command R+ 2 (July 26)

Enterprise RAG-focused release.

What shipped:

• Command R+ 2 — improved citation accuracy and multilingual retrieval.
• Embed v4 — embedding model with 8K context per chunk (up from 512).

Benchmark highlights:

• RAG citation accuracy on enterprise eval set: +11% over Command R+.
• Embed v4 MTEB: competitive with text-embedding-3-large at lower self-hosted cost.

What to do: Worth testing if you run enterprise RAG on Cohere already. Not a general-purpose model migration target for most teams.

Open-weight releases

Alibaba — Qwen 3 family (July 22)

The open-weight story of July.

What shipped:

• Qwen 3 7B, 32B, 72B — instruction-tuned, Apache 2.0 license on 7B/32B.
• Qwen 3 Coder 32B — code-specialized variant.
• Quantized GGUF releases day-one for Ollama and llama.cpp.

Benchmark highlights:

• Multilingual MMLU (29 languages): 78.4% on 72B — best open-weight multilingual score of 2026.
• Structured JSON output reliability: 94% valid-json rate on complex schemas (32B).
• SWE-bench Verified (Coder 32B): 64.1% — behind Llama 4.1 70B but strong for its size class.

What to do: If you serve non-English users and self-host, start with Qwen 3 32B on a single A100. Pair with Ollama for local dev. For English-only coding agents, Llama 4.1 70B still leads among open models. See local models vs API.

DeepSeek — V3.2 preview (July 29)

Preview, not production.

• 15% faster inference on MoE layers.
• Improved tool-call JSON formatting (fewer parse errors).
• Full release expected August. No license change.

What to do: Wait for GA. V3.1 remains the stable open reasoning pick.

Pricing changes

Trend: June's action was Mini/Flash tiers. July moved Pro/Ultra/Opus — frontier pricing is no longer frozen. Mid-tier Sonnet and GPT-5.1 held steady.

Update routing rules — see LLM Cost Optimization and Cheapest AI APIs.

What to do this month

If you build with coding agents

→ Stay on Sonnet 4.7 default. Add Opus 4.7 as an explicit escalation path for plan mode and hard refactors — not as a blanket upgrade.

If you run production API workloads

→ Re-evaluate GPT-5 Pro for tasks you routed to Sonnet purely for cost. Pro at $7.50/$30 may now win on quality-per-dollar for long generation tasks.

If you build multimodal agents

→ Prototype on Gemini 2.5 Ultra — best native video/audio understanding this month. Run safety evals before customer-facing deployment.

If you serve multilingual users

→ Benchmark Qwen 3 32B against your current API model on 50 real non-English queries. Strong candidate for self-hosted support drafts.

If you're new to model selection

→ Read June 2026 first for baseline context, then GPT vs Claude vs Gemini. Always validate on your eval set.

Coming next month (August 2026 preview)

Rumored but unconfirmed:

• DeepSeek V3.2 GA with open weights.
• Anthropic Sonnet 4.7 context extension (500K preview).
• OpenAI GPT-5.2 — incremental reasoning bump.
• Meta Llama 4.2 multimodal (image-in, text-out).

Confirmed releases are in the August 2026 changelog — same format, same section structure.

Recurring series. June 2026 · July 2026 (this page) · RSS for updates.