← Back to articles

Current AI Models as of July 2026

July 2026 turned out to be eventful: GPT-5.6, Grok 4.5, and Muse Spark 1.1 launched within a single week, and Claude Fable 5 returned after a nearly three-week pause. Below are the current models and agents covered in this article:

  • OpenAI: GPT-5.6 Sol, Terra, Luna; GPT-5.5; ChatGPT Work; ChatGPT Images 2.0; Codex
  • Anthropic: Claude Fable 5, Opus 4.8, Sonnet 5; Claude Cowork
  • Google: Gemini 3.1 Pro, Gemini 3.5 Flash; Gemini Spark; Veo 3.1
  • xAI: Grok 4.5
  • Meta: Muse Spark 1.1
  • ByteDance: Seedream 5.0 Pro
  • Alibaba: Qwen 3.7 Max (API); Qwen 3 235B-A22B, Qwen 3.5, Qwen 3.6 (open weights)
  • Open-source: Llama 4 Scout and Maverick; DeepSeek V4, R1, V3; Mistral Large 3, Small 4; GLM-5, GLM-4.7; Gemma 3, Gemma 4 26B A4B; Phi-4; Kimi K2.6

The flagship model market no longer boils down to a single winner - each family has its own strength, and the choice depends on task, budget, and context requirements.

Market overview

By mid-2026, a stable three-pole structure had formed:

  • OpenAI - GPT-5.6 (Sol, Terra, Luna) and the ChatGPT / Codex / ChatGPT Work ecosystem
  • Anthropic - Claude Fable 5, Opus 4.8, Sonnet 5, and the Claude Cowork agent
  • Google - Gemini 3.1 Pro, Gemini 3.5 Flash, and the Gemini Spark agent

Alongside them are specialized players: xAI (Grok 4.5), Meta (Muse Spark 1.1), ByteDance (Seedream 5.0 Pro for images), Alibaba (Qwen 3.7 Max). In parallel, the mature open-source segment closes the gap with closed flagships to within a few points on key benchmarks - Llama 4, Qwen 3.x, DeepSeek V4, Mistral Large 3, and GLM-5 are already production-ready with self-hosting. Competition has shifted from who is smarter to who is more efficient on a specific task at an acceptable price.

Below is the current state of the main model lines as of July 11, 2026.

OpenAI: GPT-5.6 family

On July 9, OpenAI opened GPT-5.6 to all users. The model became the default in ChatGPT after a limited preview accompanied by regulatory reviews.

Three tiers in one generation

Model Role Context Input / 1M Output / 1M
Sol Flagship: code, agents, complex tasks 1.05M $5.00 $30.00
Terra Balance of price and quality 1.05M $2.50 $15.00
Luna Fast and economical 1.05M $1.00 $6.00
GPT-5.5 Previous flagship, fallback 1M $5.00 $30.00

For input above 272K tokens, GPT-5.6 uses long-context pricing: input x2, output x1.5 (Sol: $10 / $45 per 1M).

Sol is positioned as OpenAI's best coding model: according to the Artificial Analysis Coding Agent Index, it outperforms Claude Fable 5 with lower token usage and cost. Terra and Luna cover mass scenarios - from everyday chat to background automation.

GPT-5.5 remains the fallback where proven stability is needed. Sol showed increased agentic behavior in some scenarios, so for production-critical tasks it makes sense to test both versions.

ChatGPT Work

Along with GPT-5.6, ChatGPT Work was announced - an agent for office tasks: documents, spreadsheets, presentations, simple web apps. It combines ChatGPT and Codex capabilities and competes with Claude Cowork. Integrations with Slack, Gmail, Google Drive, calendars, and CRM are supported.

Anthropic: Claude after the pause

On July 1, Anthropic brought back Claude Fable 5 - the Mythos-class flagship removed on June 12 due to US export restrictions. The model's return noticeably changed the landscape in coding and long agentic scenarios.

Key models in the lineup

Model Use case Context Input / 1M Output / 1M
Claude Fable 5 Coding, long agentic workflows 1M $10.00 $50.00
Claude Opus 4.8 Daily work, enterprise 1M $5.00 $25.00
Claude Sonnet 5 Text, instructions (intro through Aug 31) 1M $2.00 $10.00

From September 1, 2026, Claude Sonnet 5: $3.00 input / $15.00 output per 1M.

Claude Cowork - an agent for collaborative work with files and apps, OpenAI's direct response to ChatGPT Work. For teams already using Claude Code, Cowork is a natural addition to the dev stack.

Google: Gemini 3.x

Google runs the lineup in two dimensions - reasoning and price-performance.

Model Use case Context Input / 1M Output / 1M
Gemini 3.1 Pro Complex reasoning, accuracy 1M $2.00 $12.00
Gemini 3.5 Flash Price-performance, long context 2M $1.50 $9.00

Above 200K tokens, Gemini 3.1 Pro uses elevated pricing: $4.00 input / $18.00 output per 1M.

Gemini's strength is long context and multimodality. For analyzing large document sets, codebases, and mixed data (text + images), this is often the deciding factor.

Gemini Spark - Google's agent, one of the two most visible agentic products of summer 2026 alongside Claude Cowork.

xAI, Meta, ByteDance, and Alibaba

Model Use case Context Input / 1M Output / 1M
Grok 4.5 Code, real-time X context 1M $2.00 $6.00
Meta Muse Spark 1.1 Accessible mid-tier API 256K $1.25 $4.25
Qwen 3.7 Max Mid-tier reasoning (API) 1M $2.50 $7.50
Seedream 5.0 Pro Image generation - per image per image

Grok 4.5 is strong with Cursor and tasks that need up-to-date context from X and the web. Qwen 3.7 Max - closed weights via Alibaba Cloud; the open-source Qwen lineup is described below. Seedream 5.0 Pro is billed per image, not per token.

Open-source models

Open weights are a separate market layer that in 2026 is no longer a backup option for experiments. Several models match GPT-4-class on code and reasoning, and DeepSeek V4 and GLM-5.1 compete with closed flagships on SWE-Bench. Main reasons to choose open-source: self-hosting, data control, predictable cost at scale, and no vendor lock-in.

Open-source model prices below are from OpenRouter as of July 11, 2026 (hosted inference without deploying your own hardware). Self-hosting remains possible - then per-token billing does not apply.

Meta: Llama 4

Model Use case Context Input / 1M Output / 1M
Llama 4 Scout RAG, large archives 10M $0.10 $0.30
Llama 4 Maverick General-purpose 1M $0.15 $0.60

Prices - OpenRouter. With self-hosting, there is no per-token charge; cost is hardware and electricity only.

The Llama 4 Community license restricts products with over 700M MAU - for most B2B projects this is not an issue; large consumer platforms need a separate review.

Alibaba: Qwen (open weights)

Cloud Qwen 3.7 Max and the open-source lineup are different products. For self-hosting, these are relevant:

Model Use case Context Input / 1M Output / 1M
Qwen 3 235B-A22B Reasoning + coding 131K $0.46 $1.82
Qwen 3.5-397B-A17B Multilingual, science 256K $0.39 $2.45
Qwen 3.6-35B-A3B Single server, quantization 262K $0.14 $1.00

Qwen is the most common choice when you need a balance of quality, Apache 2.0 license, and support in vLLM / Ollama / llama.cpp.

DeepSeek

Model Use case Context Input / 1M Output / 1M
DeepSeek V4 Flash Coding, agents, budget 1M $0.08 $0.17
DeepSeek V4 Pro Complex reasoning 1M $0.44 $0.87
DeepSeek R1 Math, deep reasoning 164K $0.70 $2.50
DeepSeek V3.2 General-purpose fallback 131K $0.21 $0.32

On OpenRouter, DeepSeek V4 Flash is cheaper than through the direct DeepSeek API. The direct API doubles rates during peak hours (9:00-12:00 and 14:00-18:00 CST) from mid-July 2026.

Mistral AI

Model Use case Context Input / 1M Output / 1M
Mistral Large 3 Multilingual, enterprise 262K $0.50 $1.50
Mistral Small 4 Production chat + tools 262K $0.15 $0.60

Mistral wins where a permissive license without MAU cap, European data residency, and a mature deployment ecosystem matter.

Zhipu AI: GLM

Model Use case Context Input / 1M Output / 1M
GLM-5.1 Agentic coding 203K $0.97 $3.04
GLM-4.7 Coding workflows 203K $0.40 $1.75

Google and Microsoft: compact models

Model Use case Context Input / 1M Output / 1M
Gemma 3 27B Single GPU 131K $0.08 $0.16
Gemma 4 26B A4B Math + coding 262K $0.06 $0.33
Phi-4 Edge, laptop 16K $0.07 $0.14

Kimi K2.6

Model Use case Context Input / 1M Output / 1M
Kimi K2.6 Agentic coding 262K $0.66 $3.41

Kimi K2.6 (Moonshot AI) - an open-weight model focused on agentic coding and long development sessions. Often mentioned alongside DeepSeek V4 and GLM-5 as top-tier for autonomous coding agents.

Open-source flagship comparison

Model Strength Context Input / 1M Output / 1M License
Qwen 3 235B-A22B Reasoning + coding 131K $0.46 $1.82 Apache 2.0
DeepSeek V4 Flash Coding + agents 1M $0.08 $0.17 MIT
DeepSeek V4 Pro Complex reasoning 1M $0.44 $0.87 MIT
DeepSeek R1 Math + deep reasoning 164K $0.70 $2.50 MIT
Llama 4 Scout Long context 10M $0.10 $0.30 Llama 4 Community
Mistral Large 3 Multilingual + enterprise 262K $0.50 $1.50 Apache 2.0
Mistral Small 4 Production chat + tools 262K $0.15 $0.60 Apache 2.0
GLM-5.1 Agentic coding 203K $0.97 $3.04 MIT
Qwen 3.6-35B-A3B Single-server deploy 262K $0.14 $1.00 Apache 2.0
Gemma 3 27B Single GPU 131K $0.08 $0.16 Gemma
Gemma 4 26B A4B Math + coding 262K $0.06 $0.33 Apache 2.0
Phi-4 Edge / laptop 16K $0.07 $0.14 MIT
Kimi K2.6 Agentic coding 262K $0.66 $3.41 Open weights

All prices - OpenRouter, July 2026. With self-hosting, factor in GPU and electricity costs instead of per-token billing.

When to choose open-source

Open-source makes sense if:

  • data must not leave your perimeter (on-premise, private cloud)
  • you need predictable cost at scale with millions of requests per month
  • fine-tuning for your domain without API restrictions matters
  • your team already runs GPU infrastructure or uses vLLM / TGI

Closed APIs (GPT-5.6, Claude, Gemini) remain preferable when you need out-of-the-box agents (ChatGPT Work, Cowork, Spark), minimal time-to-market without DevOps, and access to the latest multimodal capabilities without deployment.

Multimodality: images and video

Text LLMs are only part of the ecosystem:

  • ChatGPT Images 2.0 - image generation with readable text on the image
  • Google Veo 3.1 - leader among video models after OpenAI shut down the Sora 2 consumer app
  • Seedream 5.0 Pro - alternative for design and marketing materials

The choice depends on whether you want a unified stack (OpenAI / Google) or the best quality in a specific modality.

How to choose a model by task

There is no universal best AI. A practical scheme:

Task What to look at first
Daily chat and documents GPT-5.6 Terra or Luna, Claude Opus 4.8
Programming and agents GPT-5.6 Sol, Claude Fable 5
Long context (code, archives) Gemini 3.5 Flash, Gemini 3.1 Pro
Text and editing Claude Sonnet 5
Budget and API automation GPT-5.6 Luna, Grok 4.5, Qwen 3.7 Max
Self-hosting and on-premise Qwen 3 235B-A22B, DeepSeek V4, Mistral Small 4, Llama 4 Scout
Coding on your own hardware DeepSeek V4, GLM-5.1, Qwen 3.6-35B-A3B
Single GPU / edge Gemma 3 27B, Phi-4, Qwen 3.6-35B-A3B (OpenRouter)
Long context (local) Llama 4 Scout, Qwen 3.6
Real-time news and X Grok 4.5
Office automation ChatGPT Work, Claude Cowork, Gemini Spark
Images / video ChatGPT Images 2.0, Veo 3.1, Seedream 5.0 Pro

For production, keep an abstraction over providers: route by task type, fallback to GPT-5.5 or Opus 4.8 on failures, and monitor token costs separately.

What changed in the last month

  1. GPT-5.6 reached GA - OpenAI regained leadership in coding benchmarks and lowered flagship cost vs competitors
  2. Claude Fable 5 is back - Anthropic restored its position in SWE and agentic scenarios
  3. Agents went mainstream - ChatGPT Work, Claude Cowork, and Gemini Spark turn LLMs from chat into work tools
  4. Price war intensified - Luna, Grok 4.5, and Muse Spark 1.1 pressure the mid-tier and split the API integration market
  5. Regulation became a factor - the Fable 5 pause showed top-model access can change quickly
  6. Open-source caught the closed frontier - DeepSeek V4, GLM-5.1, and Qwen 3.6 reached SWE-Bench levels comparable to GPT-5.5 and Opus 4.8

Summary

In July 2026, the AI market is mature and fragmented. GPT-5.6 Sol and Claude Fable 5 share first place in coding among closed APIs; Gemini 3.5 Flash leads in long context and price-performance; Claude Sonnet 5 leads in text work. In open-source, Qwen 3 235B-A22B, DeepSeek V4, and Llama 4 Scout lead - each with a narrow specialization. Agent products (ChatGPT Work, Cowork, Spark) are the next layer on top of models.

For business, the optimal strategy is not to lock into one vendor: test 2-3 models on real tasks, calculate cost per task, not just price per million tokens, and plan fallback - to the previous API generation (GPT-5.5, Opus 4.8) or self-hosted open-source (DeepSeek V4, Mistral Small 4) when data and budget require it.

Frequently asked questions

Which model is best for programming in July 2026?

Among closed APIs, GPT-5.6 Sol and Claude Fable 5 lead - both strong in coding and agentic scenarios. Sol wins on output token cost; Fable 5 on SWE-Bench Pro (~80.3%). For budget API - Grok 4.5 ($2 / $6) or GPT-5.6 Luna ($1 / $6). In open-source - DeepSeek V4 Flash on OpenRouter ($0.08 / $0.17) and GLM-5.1.

Open-source or closed API - which to choose?

Open-source (via OpenRouter or self-hosting) - if data control, predictable cost at scale, and no vendor lock-in matter. Closed API (GPT-5.6, Claude, Gemini) - if you need out-of-the-box agents (ChatGPT Work, Cowork, Spark), multimodality without DevOps, and minimal time-to-market. For many teams, a hybrid works best: closed API for product features, open-source for background automation.

How do GPT-5.6 Sol, Terra, and Luna differ?

All three are one generation with 1.05M token context but different capability levels and price. Sol ($5 / $30) - flagship for complex code and agents. Terra ($2.50 / $15) - balance for daily work. Luna ($1 / $6) - mass tasks: classification, data extraction, routine chat. ChatGPT Free and Go default to Terra.

Why was Claude Fable 5 unavailable and is it back?

On June 12, 2026, Anthropic removed Fable 5 due to US export restrictions on frontier models. On June 30 the restriction was lifted; on July 1 the model returned to the API and competing services at the same price ($10 / $50). The pause showed top-model access can change for regulatory reasons - keep a fallback (Opus 4.8, GPT-5.6 Sol).

How to save on API without losing quality?

Three practical approaches: task routing - Luna or DeepSeek V4 Flash for simple requests, Sol or Fable 5 only for complex ones; OpenRouter for open-source models - often cheaper than direct API (e.g. DeepSeek V4 Flash $0.08 vs $0.14); prompt caching and batch API - up to 50-90% savings on repeated context. Calculate cost per task, not just price per million tokens - a cheap model with long output can cost more overall.

Contacts