AI Model Changelog

Anthropic launches Claude Fable 5 — its first "Mythos-class" model

Claude Fable 5 v5

First publicly released model in Anthropic's new Mythos capability tier (a step above Opus): 1M context, $10/$50 per 1M tokens, SWE-bench Verified 95.0, GPQA Diamond 92.6. Flagged requests fall back to Opus 4.8. Suspended within days of launch (~June 12) under a US export-control directive following a jailbreak disclosure — verify current availability before relying on it.

Claude Opus 4.8 refreshes the flagship Opus tier

Claude Opus 4.8 v4.8

Flagship Opus-tier update and the safety fallback that serves requests redirected away from Fable 5. 1M context; strong agentic coding, long-document reasoning and instruction following.

Google previews Gemini 3.5 Flash

Gemini 3.5 Flash

Frontier agentic and coding performance at Flash cost — roughly 4x faster than Pro and already the default model in the Gemini app and Search. Preview availability.

xAI releases Grok 4.3

Grok 4.3 v4.3

Cost-efficient frontier reasoning with native video input and a 1M-token context; tops xAI's Artificial Analysis Intelligence Index. GPQA Diamond ~90.1 (xAI did not publish a full standard benchmark table).

Mistral Medium 3.5 replaces the Large 2.x line

Mistral Medium 3.5

New flagship-tier API model unifying chat, reasoning and coding; the official replacement for the deprecated Mistral Large 2.x family. The open-weight Mistral Large 3 (Dec 2025) remains the top open Mistral.

DeepSeek-V4 lands as an open-weight frontier MoE

DeepSeek-V4

~1.6T-total / ~49B-active mixture-of-experts with 1M context, MIT-licensed open weights plus API. An optional thinking mode absorbs the old R1 reasoning role. Headline benchmarks are vendor-reported (no technical report yet); the arXiv-verified V3.2 baseline is GPQA 82.4 / SWE-bench Verified 73.1.

OpenAI ships GPT-5.5

GPT-5.5

First ground-up retrained base since GPT-4.5, with unified built-in reasoning (the standalone o-series branding is retired) and a ~1.1M-token context. GPQA Diamond 93.5, SWE-bench Verified 82.6. Tops academic indices but trails Claude/Gemini on LMArena human-preference voting.

Meta debuts Muse Spark — and retires the Llama brand

Muse Spark

Meta Superintelligence Labs' first closed-weights frontier model: native multimodal reasoning with thought compression and parallel multi-agent orchestration; powers Meta AI. Ends the Llama open-weight lineage as Meta's flagship. GPQA 89.5, SWE-bench Verified 77.4.

Meta Releases Llama 3.2 405B — Largest Open Source Model

Llama 3.2 v405B

Meta releases Llama 3.2 with a 405B parameter model, the largest open-source LLM. Includes vision variants at 11B and 90B.

• ✅405B parameter model — largest open-source LLM
• ✅Vision variants: 11B and 90B with image understanding
• ✅128K token context window
• ✅Competitive with GPT-4o on many benchmarks

Gemini 3.1 Flash Image Preview

Gemini 3.1 Flash Image Preview

Preview release of Gemini 3.1 Flash with native image generation capabilities alongside text.

• ✅Native image generation within conversations
• ✅Text + image multimodal output
• ✅Preview access via AI Studio

GPT-5.4 Mini & Nano — Speed and Efficiency

GPT-5.4 Mini

Smaller, faster variants of GPT-5.4 optimized for latency-sensitive and cost-sensitive applications.

• ✅GPT-5.4 Mini: 128K context, balanced performance
• ✅GPT-5.4 Nano: 32K context, ultra-fast responses
• ✅10x cheaper than GPT-5.4 Thinking

OpenAI Unveils GPT-5.4 Thinking System

GPT-5.4 Thinking

GPT-5.4 Thinking introduces deliberative reasoning with chain-of-thought transparency. Tops benchmarks in math, coding, and complex analysis.

• ✅Deliberative reasoning with transparent chain-of-thought
• ✅256K token context window
• ✅Improved mathematical reasoning
• ✅Enhanced code generation

OpenAI Codex Security — AI Agent for Automated Patching

Codex

Codex Security automates vulnerability detection and patching. Integrates with GitHub Actions for CI/CD security scanning.

• ✅Automated vulnerability detection
• ✅AI-generated patches for common CVEs
• ✅GitHub Actions integration
• ✅Support for Python, JavaScript, Go, Rust

Claude Code 2.1.76 — Dispatch, Channels, MCP Elicitation

Claude Code v2.1.76

Major Claude Code update with background task dispatch, agent-to-agent channels, and MCP elicitation support.

• ✅Dispatch for background task orchestration
• ✅Channels for structured agent communication
• ✅MCP elicitation support
• ✅New hooks for agent lifecycle events

Claude Opus 4.6 with 1M Token Context Window

Claude Opus 4.6 v4.6

Anthropic releases Claude Opus 4.6 with a 1 million token context window, 76% accuracy on needle-retrieval benchmarks, and 40% improved SWE-bench performance.

• ✅1M token context window (5x increase from 200K)
• ✅76% accuracy on needle-retrieval benchmark (up from 18.5%)
• ✅40% improvement on SWE-bench coding benchmarks
• ✅Flat pricing at $5/$25 per 1M tokens (no multiplier)

Claude Sonnet 4.6 — Twice the Speed, Same Price

Claude Sonnet 4.6 v4.6

Claude Sonnet 4.6 delivers 2x faster output than Sonnet 4.5 with no price increase. Improved instruction following and code generation quality.

• ✅2x faster output speed
• ✅Improved instruction following
• ✅Better code generation quality
• ✅200K token context window

GPT-3.5 Turbo Deprecated — End of Life July 2026

GPT-3.5 Turbo

OpenAI announces GPT-3.5 Turbo will be fully deprecated in July 2026. Users should migrate to GPT-4o Mini.

• 💥GPT-3.5 Turbo deprecated — EOL July 2026
• 💥All API calls will return errors after July 14, 2026
• ✅Migration path: use GPT-4o Mini (cheaper and better)

xAI Launches Grok 3 with Real-Time Data

Grok 3 v3.0

Grok 3 features real-time X/Twitter data access, unfiltered responses, and improved reasoning. Available to X Premium+ subscribers.

• ✅Real-time X/Twitter data integration
• ✅128K token context window
• ✅Unfiltered response mode
• ✅X Premium+ exclusive

Gemini 2.5 Flash — Speed Champion

Gemini 2.5 Flash

Fastest model in the Gemini family with 1M context and thinking mode at a fraction of Pro pricing.

• ✅Ultra-fast inference
• ✅1M token context
• ✅Thinking mode support
• ✅15 RPM free tier in AI Studio

Gemini 2.5 Pro — 1M Context with Thinking Mode

Gemini 2.5 Pro

Google releases Gemini 2.5 Pro with 1M token context, extended thinking mode for complex reasoning, and Google Search grounding.

• ✅1M token context window
• ✅Extended thinking mode for step-by-step reasoning
• ✅Google Search grounding for real-time information
• ✅Native multimodal: text, image, audio, video, code

GPT-4o Price Reduction — 50% Cheaper Input

GPT-4o

OpenAI reduces GPT-4o input pricing by 50% to stay competitive with Gemini Flash and DeepSeek.

• 📝Input price reduced from $5 to $2.50 per 1M tokens
• 📝Output price unchanged at $10 per 1M tokens

Claude 3 Opus & Sonnet Deprecated

Claude 3 Opus

Anthropic deprecates Claude 3 Opus and Claude 3 Sonnet models. Users should migrate to Claude 4.6 family.

• 💥Claude 3 Opus deprecated
• 💥Claude 3 Sonnet deprecated
• ✅Migration path: Claude Sonnet 4.6 (faster and cheaper)

Mistral Large 2 — European AI Powerhouse

Mistral Large 2

Mistral releases Large 2 with 128K context, multilingual excellence, and competitive coding benchmarks. Leading European AI model.

• ✅128K token context
• ✅Excellent multilingual performance (23 languages)
• ✅Strong coding benchmarks
• ✅Available on La Plateforme and major clouds

DeepSeek R1 — Open Source Reasoning Model

DeepSeek R1

DeepSeek releases R1, an open-source reasoning model with strong performance at extremely low cost. Raises concerns about Chinese data governance.

• ✅Open source with MIT license
• ✅Strong reasoning capabilities
• ✅128K token context
• ✅Extremely affordable API pricing