Model Release Changelog — June 2026

Recurring format: This is the first edition of AIKnowHub's monthly model release changelog. Each month we cover frontier closed-model releases, open-weight drops, pricing changes, and a practical recommendation. Bookmark this series — or check /learn/model-release-changelog-june-2026 for the latest edition.

June 2026 at a glance

Frontier closed models

Anthropic — Claude 4.7 Sonnet (June 3)

The biggest coding-agent release of the month.

What shipped:

• Claude 4.7 Sonnet — 15% faster tool-use loops, improved multi-file edit accuracy.
• Claude 4.7 Opus — incremental quality bump on hardest reasoning tasks. Still the flagship.
• Claude Code and Cursor both updated default backends within 48 hours.

Benchmark highlights:

• SWE-bench Verified: 72.1% (up from 68.4% on 4.5 Sonnet).
• Multi-file edit accuracy: +8% on internal Anthropic eval.
• Agent loop latency: 15% reduction (measured tool-call → response time).

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

What to do: If you use Claude Code or Cursor with Claude, you likely already have 4.7. Run your coding eval set — upgrade routing if agent tasks improved. No migration needed for API users; it's the same endpoint.

OpenAI — GPT-5.1 (June 12)

Incremental but meaningful — especially on cost.

What shipped:

• GPT-5.1 — reasoning mode 2x faster than GPT-5. Quality bump on math and planning tasks.
• GPT-5 Mini price cut — 40% reduction to $0.15/$0.60 per 1M tokens (input/output).
• ChatGPT image generation quality improvement — better text rendering in images.
• Structured outputs schema mode now supports nested arrays (API change, all GPT-5 variants).

Benchmark highlights:

• MATH benchmark: 94.2% (up from 91.8%).
• Reasoning mode latency: 2x faster (median 8s → 4s on AIME-level problems).
• GPT-5 Mini MMLU: within 2 points of full GPT-5.

Pricing: GPT-5.1 same as GPT-5 ($2.50/$10 per 1M). Mini cut 40%. See Cheapest AI APIs for updated comparison.

What to do: Switch classification and extraction workloads to GPT-5 Mini if you're on Haiku or older Mini variants — 40% cheaper with near-frontier quality. Re-run LLM cost optimization routing rules.

Google — Gemini 2.5 Pro 2M context (June 18)

Long-context goes GA.

What shipped:

• Gemini 2.5 Pro — 2M token context window generally available (was preview).
• Gemini 2.5 Flash — speed-optimized variant for high-throughput tasks.
• Google Workspace side panel — Gemini in Docs, Sheets, Slides rollout complete for Workspace Business plans.

Benchmark highlights:

• Long-context retrieval (needle-in-haystack at 1M tokens): 99.1% accuracy.
• Coding (HumanEval): 92.4% — competitive but still behind Claude 4.7 on multi-file tasks.
• Price per 1M tokens: Gemini 2.5 Pro $1.25/$5.00 — still the cheapest frontier option.

What to do: Use Gemini 2.5 Pro for long-document workflows (slides from document, contract review, research synthesis). For coding agents, Claude 4.7 still leads — use Gemini for doc processing. See GPT vs Claude vs Gemini.

Open-weight releases

Meta — Llama 4.1 70B (June 20)

The open-weight story of the month.

What shipped:

• Llama 4.1 70B — instruction-tuned, code-focused variant.
• Llama 4.1 8B — edge/mobile variant, runs on M-series Macs with 16GB RAM.
• License: unchanged (Meta custom license, commercial use allowed under thresholds).

Benchmark highlights:

• SWE-bench Verified: 67.3% — within 5 points of Claude 4.7 Sonnet (72.1%).
• HumanEval: 91.8%.
• Runs on 2x A100 80GB or 1x H100 with quantization.

What to do: If you self-host and your primary workload is coding, test Llama 4.1 70B — it may replace API calls for dev-time agent tasks. For production customer-facing features, API models still have better safety tuning and structured output reliability. See local models vs API (when published).

Mistral — Codestral 2 (June 25)

What shipped:

• Codestral 2 — 22B parameter code completion and fill-in-the-middle model.
• Mistral Large 3.1 — minor quality bump on reasoning, no benchmark changes claimed.

Benchmark highlights:

• Codestral 2 HumanEval: 93.1% — best open code-completion model.
• Fill-in-the-middle accuracy: +12% over Codestral 1.

What to do: Codestral 2 is excellent for IDE inline completions (pair with Continue or self-hosted). Not a replacement for full agent loops — use Claude/GPT for multi-file tasks.

DeepSeek — V3.1 patch (June 28)

Minor release.

• Bug fixes in tool-calling format.
• 10% inference speed improvement on H100.
• No benchmark changes. DeepSeek V3 remains the best open-weight reasoning model.

Pricing changes

Trend: Mini/Flash/Haiku tiers are in a price war. Mid-tier (Sonnet, GPT-5, Gemini Pro) holding steady. Opus and GPT-5 Pro unchanged.

Update your model routing — see LLM Cost Optimization.

What to do this month

If you build with coding agents

→ You're probably already on Claude 4.7 via Cursor or Claude Code. Run your eval set. No action if results are stable.

If you run production API workloads

→ Re-route classification/extraction to GPT-5 Mini (40% cheaper). Keep Sonnet/Opus for generation quality. Update cost dashboards.

If you process long documents

→ Test Gemini 2.5 Pro 2M context for slide extraction and contract review. Compare quality against your current model on 5 real docs.

If you self-host

→ Download Llama 4.1 70B. Benchmark on your coding tasks. If within acceptable quality, shift dev-time agent calls off API.

If you're choosing a model for the first time

→ Read GPT vs Claude vs Gemini and Best AI Model for Coding. Don't pick based on this changelog alone — pick based on your eval set.

Coming next month (July 2026 preview)

Rumored but unconfirmed:

• Anthropic Opus 4.7 public release (currently limited access).
• OpenAI GPT-5 Pro price reduction (following Mini cut pattern).
• Google Gemini 2.5 Ultra announcement.
• Qwen 3 open-weight family from Alibaba.

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

This is a recurring series. Each edition follows this template: at-a-glance table, frontier closed models, open-weight releases, pricing changes, and actionable recommendations. Subscribe via RSS or bookmark this page for monthly updates.