AI Model Releases — Week of July 9, 2026

Executive Summary

The July 2–9 release window brought an unusually broad set of model launches. OpenAI moved the full GPT‑5.6 family from limited preview to general availability; Meta launched Muse Spark 1.1 with a public model API; xAI shipped Grok 4.5; and OpenAI separately introduced full-duplex GPT‑Live voice models. At the same time, Tencent, Meituan, Mistral, Poolside, and Sber expanded the open-weight frontier with models aimed at agents, coding, formal verification, and long-context work.

The defining shift was not a single benchmark win. It was the convergence of four product patterns:

1. Agent orchestration is becoming a model-level capability. GPT‑5.6 Ultra and Muse Spark 1.1 can coordinate parallel subagents rather than only emit tool calls sequentially.
2. One-million-token context is moving into mainstream flagship tiers. GPT‑5.6 exposes a 1.05M-token window, while Muse Spark 1.1 actively manages a 1M-token working context.
3. Open models are specializing. Leanstral 1.5 targets Lean 4 proof engineering, Robostral Navigate targets embodied navigation, and Laguna XS 2.1 targets local agentic coding.
4. Efficiency is now part of the intelligence claim. Vendors increasingly report tokens, latency, active parameters, and cost alongside raw benchmark scores.

Coverage is limited to releases and material public-weight/API availability changes announced from July 2 through July 9, 2026. Benchmark figures are vendor-reported unless explicitly described as independent.

---

1. ⚫ OpenAI — GPT‑5.6 Sol, Terra, and Luna

Released: July 9, 2026
Category: Frontier reasoning, coding, agents, and knowledge work
Status: General availability

OpenAI moved its three-model GPT‑5.6 family from limited preview to general availability. The family replaces a single flagship tier with three persistent capability classes: Sol for maximum capability, Terra for balanced production workloads, and Luna for cost-sensitive, high-volume use.

Key Specs

All three accept text and images, support function calling, web search, file search, and computer use, and expose reasoning levels from none through max. The published knowledge cutoff is February 16, 2026.

Benchmarks and Capabilities

• Agents’ Last Exam: GPT‑5.6 Sol scored 53.6, 13.1 points above Claude Fable 5 under OpenAI's reported comparison.
• Artificial Analysis Coding Agent Index: Sol with max reasoning scored 80, which OpenAI describes as a new high.
• BrowseComp: Sol reached 92.2%.
• OSWorld 2.0: Sol reached 62.6%.
• Terminal-Bench 2.1: OpenAI's launch chart reports 88.8% for Sol and 91.9% for the multi-agent Ultra configuration.

The most consequential feature is Ultra mode. Rather than simply giving one model a larger reasoning budget, Ultra coordinates four agents in parallel by default. OpenAI positions this as a latency-capability trade: more total tokens, but faster time-to-result on work that can be split into independent branches. Developers can build similar workflows through the Responses API multi-agent beta.

GPT‑5.6 also adds Programmatic Tool Calling, allowing the model to write small programs that filter tool output, preserve only relevant state, and decide the next action without returning every intermediate result through another full model round trip.

Availability

• ChatGPT and ChatGPT Work surfaces
• OpenAI Responses API and client SDKs
• Model IDs: gpt-5.6-sol, gpt-5.6-terra, and gpt-5.6-luna
• Batch and Flex pricing are 50% below standard token rates

Sources: OpenAI launch, model catalog, API pricing

---

2. 🔵 Meta — Muse Spark 1.1

Released: July 9, 2026
Category: Multimodal reasoning and agent orchestration
Status: Public API preview

Muse Spark 1.1 is Meta Superintelligence Labs' second release in the Muse Spark family and the first available to outside developers through the new Meta Model API. It is a multimodal reasoning model built around tool use, computer use, coding, and multi-agent execution.

Key Capabilities

• 1M-token managed context: the model retrieves earlier actions and compacts its working state while preserving steps needed later.
• Parallel subagents: it can act as a main agent that delegates work or as a constrained subagent that escalates when necessary.
• Computer use: it chooses between scripts, direct interface actions, and batches of actions depending on the task.
• Multimodal workflows: it can inspect images, audio, video, and documents while operating external tools.
• Agent compatibility: Meta says it zero-shot generalizes to new native tools, MCP servers, and custom skills.

Meta's launch article emphasizes end-to-end agent performance rather than a single public leaderboard number. The accompanying evaluation report covers coding, reasoning, health, multimodal, agent, and safety evaluations, but the results depend on Meta's own harness and comparison methodology.

Safety Note

Meta's evaluation report says that, before mitigations, it could not rule out Muse Spark 1.1 reaching the company's “high risk” capability threshold in chemical/biological and cybersecurity domains. Meta reports that deployment safeguards reduce residual risk to “moderate or lower,” which is why it proceeded with release.

Availability

• Public preview through the Meta Model API
• “Thinking” mode in the Meta AI app and on meta.ai
• Pricing was not published in the launch materials reviewed for this report

Sources: Meta launch, Muse Spark 1.1 evaluation report, evaluation methodology

---

3. ⚪ xAI / SpaceXAI — Grok 4.5

Released: July 8, 2026
Category: Coding, agentic work, and knowledge work
Status: API and product availability

Grok 4.5 is xAI's new frontier model for coding, agents, and professional knowledge work. Its developer surface combines a large context window with built-in search and code execution tools.

Key Specs

The launch frames token efficiency as a primary advantage, claiming roughly twice the token efficiency of the previous comparison baseline for agentic work. Because the public announcement page was not consistently retrievable during research, this report does not reproduce its benchmark chart as a cross-vendor score table.

Availability

• xAI API through Responses and Chat Completions interfaces
• Default model in Grok Build
• Available in Cursor plans
• EU API availability was still pending at launch

Sources: Grok 4.5 announcement, model documentation, developer guide, pricing

---

4. ⚫ OpenAI — GPT‑Live‑1 and GPT‑Live‑1 mini

Released: July 8, 2026
Category: Full-duplex conversational voice
Status: Rolling out in ChatGPT; API planned

GPT‑Live replaces rigid turn-based voice interaction with a full-duplex architecture that listens and speaks at the same time. The model continuously decides whether to speak, pause, keep listening, interrupt, or invoke a tool.

Architecture

GPT‑Live separates two concerns:

1. Continuous interaction stays with the low-latency voice model.
2. Deeper work—search, reasoning, or agentic execution—is delegated to a frontier model in the background.

At launch, GPT‑Live uses GPT‑5.5 for delegated reasoning. OpenAI says the backing frontier model can change over time without replacing the real-time interaction layer.

Evaluations

OpenAI reports that GPT‑Live‑1 and GPT‑Live‑1 mini were preferred over the previous Advanced Voice Mode in matched 5–10 minute conversations covering turn-taking, interruptions, flow, and perceived naturalness. GPT‑Live‑1 also improved on GPQA, BrowseComp, and an internal telecom voice-agent evaluation, though the release article does not expose a complete numeric comparison table in text.

Availability

• Global rollout in ChatGPT Voice
• GPT‑Live‑1 for higher-capability use and GPT‑Live‑1 mini for the lower-cost tier
• API availability announced as “coming soon”

Sources: OpenAI launch, system card

---

5. 🟠 Tencent — Hy3

Released: July 6, 2026
Category: Open-weight reasoning and agent model
License: Apache 2.0

Hy3 is the production release that follows Tencent's April preview. It is a 295B-parameter Mixture-of-Experts model with 21B parameters activated per token and a separate 3.8B-parameter multi-token-prediction layer.

Architecture and Specs

Benchmarks

• SWE-bench Verified: 78.0%
• GPQA Diamond: 90.4%
• Internal expert workflow evaluation: 2.67/4, versus 2.51/4 for GLM‑5.1

Tencent also reports that post-training reduced its internal hallucination rate from 12.5% to 5.4% and multi-turn issue rate from 17.4% to 7.9%. These are useful product signals, but they are internal evaluations and should not be treated as independent measurements.

Availability

• Open weights on Hugging Face and other Chinese model hubs
• BF16 and FP8 checkpoints
• vLLM and SGLang recipes with speculative decoding
• Tencent recommends eight high-memory GPUs for self-hosting the full model

Sources: Tencent announcement, official model card, GitHub repository

---

6. 🟡 Meituan — LongCat‑2.0

Public weights: Released during the July 2–9 window
Category: Open-weight long-context coding and agent model
License: MIT

LongCat‑2.0 is one of the largest openly downloadable MoE releases of the week: 1.6T total parameters with roughly 48B activated per token. Meituan says both training and large-scale deployment ran entirely on AI ASIC superpods rather than a conventional NVIDIA GPU cluster.

Architecture

• Pretrained on more than 35T tokens
• Trained on hundreds of billions of tokens of 1M-context data
• Uses LongCat Sparse Attention, combining streaming-aware, cross-layer, and hierarchical token indexing
• Includes 135B n-gram embedding parameters
• Supports multi-token prediction for speculative decoding
• Integrated with agent harnesses including Claude Code, OpenClaw, and Hermes

Benchmarks

Meituan notes that most scores were measured in-house under one harness, while selected competitor scores were taken from official reports. Cross-vendor comparisons therefore remain directional.

Availability

• BF16 and INT8 weights on Hugging Face
• GPU deployment through SGLang
• NPU deployment through SGLang-FluentLLM
• Official chat interface at longcat.ai

Sources: official model card, GitHub repository

---

7. 🟣 Cognition — SWE‑1.7

Released: July 8, 2026
Category: Proprietary software-engineering agent model
Availability: Devin

SWE‑1.7 is Cognition's most capable software-engineering model to date. It starts from a Kimi K2.7 base and applies additional reinforcement learning focused on long-horizon asynchronous work.

Benchmarks

Training Notes

Cognition attributes the gains to four changes:

• entropy-preserving and numerically stable RL
• rollout generation across four datacenters on three continents
• stricter task filtering and reward-hacking defenses
• learned self-compaction for tasks that exceed the raw context window

The model is served through Cerebras at a claimed 1,000 tokens per second, making serving speed part of the product proposition rather than a separate infrastructure choice.

Source: Cognition launch and training report

---

8. 🔴 Mistral — Robostral Navigate

Released: July 8, 2026
Category: Embodied navigation
Status: Research and enterprise access

Robostral Navigate is Mistral's first embodied-navigation model. The 8B model turns one RGB camera stream plus a natural-language instruction into movement decisions for wheeled, legged, and flying robots.

Key Results

Mistral reports that the unseen score is 9.7 points above the best prior single-camera approach and 4.5 points above the best compared multi-sensor system. Unlike those systems, Robostral does not require depth sensors, LiDAR, or multiple cameras.

Training and Control

• Built in-house rather than initialized from an external open VLM
• Trained on roughly 400,000 simulated trajectories across 6,000 scenes
• Predicts image coordinates and desired orientation through pointing-based navigation
• Uses local-coordinate displacement when the destination is outside the camera view
• Prefix-caching reduces supervised-training tokens by 22×
• Online reinforcement learning improved success by 3.2 points

Source: Mistral launch

---

9. 🔵 Meta — Muse Image and Muse Video

Released: July 7, 2026
Category: Agentic image and video generation

Meta launched Muse Image and previewed Muse Video, the first media-generation models from Meta Superintelligence Labs.

Muse Image is not presented as a one-pass prompt-to-image system. It can invoke web search and code execution, generate supporting assets such as plots or QR codes, inspect its own output, and iterate. It can also coordinate with Muse Spark to build richer artifacts such as websites, games, and animated media.

Muse Video shares the media pretraining base and adds native audio. Meta describes it as a forthcoming creator and Meta AI surface rather than a generally available model at launch.

Availability

• Muse Image: Meta AI app, meta.ai, Instagram Stories in the US, and WhatsApp in selected countries
• Muse Video: preview; broader availability coming later
• No open weights or standalone API pricing were announced

Source: Meta launch

---

10. 🔴 Mistral — Leanstral 1.5

Released: July 2, 2026
Category: Formal verification and Lean 4 proof engineering
License: Apache 2.0

Leanstral 1.5 is a 119B-total / 6B-active MoE model specialized for theorem proving and code verification in Lean 4. Its training combines mid-training, supervised fine-tuning, and online RL through Mistral's CISPO algorithm.

Benchmarks

The most notable result is its test-time scaling. On PutnamBench, performance continued to rise as the budget increased from 50K to 4M tokens per attempt. Mistral reports an average cost of roughly $4 per problem, far below the highest-compute comparison systems.

Real-World Verification

Mistral used a pipeline that translates Rust into Lean, generates candidate correctness properties, and tries to prove or disprove them. Across 57 repositories it identified 11 genuine bugs, five of which were reportedly previously unreported.

Availability

• Open weights on Hugging Face
• Free beta API as leanstral-1-5
• Integrated into Mistral Vibe

Source: Mistral launch

---

11. 🟦 Poolside — Laguna XS 2.1

Released: July 2, 2026
Category: Local agentic coding
License: OpenMDW 1.1

Laguna XS 2.1 is a 33B-total / 3B-active MoE model designed to run agentic coding workloads on local hardware. It keeps the previous XS.2 architecture while improving multilingual software-engineering and terminal performance.

Benchmarks

The model supports a 256K context window through Poolside's API. Poolside released BF16, FP8, INT4, and NVFP4 checkpoints plus draft “DFlash” speculator models that the company says double local generation throughput in its tests.

Pricing and Availability

• Input: $0.10 / 1M tokens
• Output: $0.20 / 1M tokens
• Cache reads: $0.05 / 1M tokens
• API, OpenRouter, Hugging Face, Ollama, vLLM, SGLang, and TensorRT-LLM

Sources: Poolside launch, model card

---

12. 🟢 Sber — GigaChat 3.5 Ultra

Released: July 6, 2026
Category: Russian-language flagship, coding, and agents
Status: Free assistant access and downloadable weights

GigaChat 3.5 Ultra is Sber's new flagship model for coding, mathematics, long-document analysis, and autonomous workflows. It uses a proprietary MoE architecture with linear attention, allowing it to accumulate a compressed representation of earlier context instead of repeatedly attending over the entire prefix.

Sber reports that the model:

• processes long texts up to four times faster than its predecessor
• is nearly half the previous model's size
• performs better on programming, mathematics, multi-step work, and Russian dialogue
• approaches DeepSeek 3.2 on selected internal metrics

The announcement does not provide a reproducible benchmark table or detailed parameter count, so the claims should be treated as preliminary until the full model card and independent evaluations are available.

Availability

• Free in the GigaChat assistant
• Downloadable for developers to integrate into services and agents
• Self-hosting positioned for less expensive hardware than the previous Ultra generation

Source: Sber announcement

---

Notable Ecosystem Update — NVIDIA Isaac GR00T 1.7 in LeRobot

NVIDIA and Hugging Face brought the open Isaac GR00T 1.7 reasoning vision-language-action model and the Isaac Teleop data-collection framework into LeRobot. This is primarily an ecosystem and distribution update rather than a new frontier-model launch, but it matters because it standardizes post-training, evaluation, and deployment for humanoid robotics inside a widely used open-source stack.

NVIDIA also said Cosmos 3 integration is planned next. The combined workflow is intended to connect teleoperation data collection, simulation, robot-policy training, evaluation, and deployment.

Source: NVIDIA and Hugging Face announcement

---

Benchmark Comparison Summary

Agentic and Coding Benchmarks

These scores are not directly interchangeable. Model vendors use different agent harnesses, reasoning budgets, patch sets, step limits, and benchmark versions. Terminal-Bench 2.0 and 2.1 are shown together only for orientation, not as a strict ranking.

Specialized Benchmarks

---

Pricing and Openness Overview

---

Week Themes

1. Multi-agent orchestration moved into flagship products. GPT‑5.6 Ultra and Muse Spark 1.1 treat decomposition and parallel execution as native model behavior, not merely an application-layer framework.
2. Context windows crossed the million-token threshold without becoming the only story. Vendors now pair long context with compaction, retrieval, and state management because raw capacity alone does not preserve task coherence.
3. The open-weight frontier is increasingly MoE. Hy3, LongCat‑2.0, Leanstral 1.5, Laguna XS 2.1, and GigaChat 3.5 Ultra all use sparse or specialist architectures to reduce active compute relative to total capacity.
4. Specialization is producing credible product categories. Formal proof engineering, robot navigation, local coding agents, and full-duplex voice are now first-class release targets.
5. Serving economics became part of model design. Active parameters, cached-input pricing, speculative decoding, quantized checkpoints, and tokens per second are central release metrics.
6. Benchmark comparability is getting worse. The leading scores increasingly depend on proprietary harnesses, multi-agent settings, long reasoning budgets, and patched environments.
7. Open distribution is broadening beyond GPUs. LongCat‑2.0's alternative-ASIC training and NPU deployment, plus GR00T's LeRobot integration, show vendors optimizing for ecosystems rather than a single inference stack.

---

What to Watch Next

• Whether GPT‑5.6 Ultra's multi-agent gains survive independent cost-normalized evaluation
• Meta Model API pricing, rate limits, and production terms for Muse Spark 1.1
• Independent coding and long-context evaluations for Hy3 and LongCat‑2.0
• Real-world navigation tests for Robostral outside Mistral's simulated training distribution
• API availability and pricing for GPT‑Live
• Whether formal-verification agents such as Leanstral produce repeatable bug-finding gains across larger software corpora
• How quickly open-weight models receive stable quantization, serving, and agent-harness support

---

Sources and Research Method

Primary Release Sources

• OpenAI — GPT‑5.6
• OpenAI — GPT‑Live
• Meta — Muse Spark 1.1
• Meta — Muse Image and Muse Video
• xAI — Grok 4.5
• Tencent — Hy3
• Cognition — SWE‑1.7
• Mistral — Robostral Navigate
• Mistral — Leanstral 1.5
• Poolside — Laguna XS 2.1
• Sber — GigaChat 3.5 Ultra
• NVIDIA and Hugging Face — GR00T 1.7 in LeRobot

Technical Documentation and Model Cards

• OpenAI model catalog
• OpenAI API pricing
• Grok 4.5 model documentation
• Tencent Hy3 model card
• Meituan LongCat‑2.0 model card
• Poolside Laguna XS 2.1 model card

Graph Trace

Research used parallel retrieval branches for official vendor announcements, dated news coverage, open-weight model cards, and specialist research releases. The first pass returned 118 unique URLs. Results were deduplicated and reranked for the July 2–9 window; outdated releases, reposts, product-only updates, and claims without a recoverable primary source were pruned. Full content was fetched only for the surviving official announcements and model documentation listed above.