Claude Sonnet 5 launches as Google expands image and video models

Anthropic launched Claude Sonnet 5 on June 30, 2026. It is now the default model for free and Claude Pro users and is also available through Claude Code and the Claude Platform. Through August 31, it costs $2 per million input tokens. Anthropic positions Sonnet 5 as a cheaper alternative to its Opus line and highlights improvements in coding, reasoning, and tool use.¹²

The release gives companies running agents for multistep development work another option besides a top-tier model. Whether it handles a particular workflow well still needs to be tested against that workflow’s tasks and costs. OpenRouter lists a context window of up to one million tokens for Sonnet 5; that figure is not confirmed by Anthropic in the provided launch information.¹³

Anthropic restored global access to Claude Fable 5 and Claude Mythos 5 starting July 1, 2026, after U.S. export controls on both models were lifted.¹ Access is returning through Claude.ai, Claude Platform, Claude Code, and Claude Cowork.¹ Fable 5 is also available through the Claude API and consumption-based Enterprise plans.²

The models had been suspended since export controls took effect on June 12.¹ Anthropic describes Fable 5 as a “Mythos-class” model suitable for general use.² Media coverage also describes tighter restrictions on cybersecurity-related uses.³⁴ For businesses, restored access is only part of the decision: they need to check the usage rules and billing terms for their access route before putting the model into a workflow.¹²

Google introduced two generative media models on June 30, 2026. Nano Banana 2 Lite is, according to Google, its fastest and most cost-efficient Nano Banana model for generating and editing images. Gemini Omni Flash generates video and supports editing through follow-up instructions, with text, images, and video as inputs. The video model is available to developers in public preview through Google AI Studio, the Gemini API, and the Gemini Enterprise Agent Platform.¹²

The releases target applications that create and revise images or videos at scale. Google positions Nano Banana 2 Lite as a replacement for its older Nano Banana model.²³ For businesses, the distinction matters: the image model is intended to reduce cost and waiting time, while Gemini Omni Flash enables new video workflows but remains a preview.¹²

Zhipu AI, also known as Z.ai, introduced GLM-5.2 in mid-June 2026 for coding and long-running agent tasks.¹² Reports put the open-weight model’s context window at one million tokens, and the weights are available under the MIT license.¹³ Companies can download and self-host the model for commercial use.³

The large context window is intended for tasks that require a model to work across extensive code and longer task histories.¹ Z.ai points to strong coding benchmark results and reports narrow gaps behind Claude Opus 4.8 on two agent-task tests.¹ Those scores alone do not establish how reliably the model will perform in a particular development environment. For teams with their own infrastructure, the release adds another commercially usable option beyond models accessible only through an API.³

Anthropic added five features to Claude Managed Agents on June 30, 2026.¹ A session can now override selected settings from a stored agent without permanently changing the agent definition.¹² Teams can also restrict which credentials from a vault are injected into a session.¹²

The changes matter for teams using one agent across workflows with different instructions or permissions: they can set exceptions for a single run instead of creating separate agents.¹² Credential scoping is intended to keep a session from receiving access to an entire vault by default.¹² This is an update to an existing platform, not the launch of a new product.¹² Companies should still check the permissions granted to their agents and connected tools.

Meta is working on a cloud business called Meta Compute, according to coverage published July 1. It would let outside customers access models hosted on Meta’s infrastructure or rent GPU compute directly.¹² Meta has not announced a public offering; the plans describe a potential service rather than a cloud product customers can already use.¹²

The move would let Meta sell infrastructure capacity to other companies instead of using it solely for its own AI offerings.¹² That could put it alongside established cloud providers and companies specializing in GPU rentals.¹ Whether the plans become a product, and when, remains unclear.¹² Prospective customers still lack the information needed to compare its costs, availability, and capabilities with existing services.¹²

DeepSeek introduced DSpark in late June 2026 as a speculative decoding method for DeepSeek-V4 Flash and DeepSeek-V4 Pro. A smaller drafting step proposes upcoming tokens for the main model to check. DSpark adjusts that step based on confidence and hardware load. The aim is faster output without replacing the underlying language model.¹

DeepSeek reports throughput gains of about 51% to 400% in certain serving scenarios. Throughput measures how much a system processes overall; the figure does not mean every individual response is that much faster.¹² The reported speedups for users on DeepSeek’s production systems are narrower and were measured at matched throughput.¹ Companies should therefore assess response time and capacity under concurrent load separately. Results should not be assumed to carry over unchanged to other hardware or workloads.¹²

NVIDIA introduced Nemotron-Labs-TwoTower on July 1, 2026, as a research approach to faster text generation. The system uses a pretrained Nemotron model: one component maintains context while another generates tokens in parallel. NVIDIA reports text generation that is 2.42 times faster than the autoregressive baseline, while TwoTower retains 98.7% of its measured quality.¹²

The diffusion architecture builds on a frozen autoregressive model rather than training an entirely new language model. It explores a way to speed up token output that would otherwise proceed sequentially. The speed and quality figures come from NVIDIA’s own tests and should not be assumed to hold across other workloads. The quality figure is an aggregate evaluation, not a claim that the models produce identical answers.¹²

Hao AI Lab at UC San Diego has introduced JetSpec, a speculative-decoding method for language models.¹ In tests of Qwen3-8B on H100 GPUs, the team reports a speedup of up to 9.64× on MATH-500. Its evaluation says output quality was preserved.²³ The research is publicly available as a paper.²

JetSpec uses “causal parallel tree drafting”: Multiple possible continuations are drafted and then checked by the target model. The approach aims to improve draft quality while controlling the work required to produce drafts.⁴ The performance figures come from the research team’s tests. They do not establish whether similar gains apply to other models, prompts, batch sizes, or production systems.²⁴ Operators also need to determine whether faster generation reduces cost per request.

Meta introduced Brain2Qwerty v2 on June 29, 2026. The system reconstructs typed sentences from brain signals recorded with magnetoencephalography (MEG), without an implant.¹ Its end-to-end deep learning model achieved an average word accuracy of 61% across participants.¹² Meta is releasing the training code for both Brain2Qwerty v1 and v2.¹

The system does not read freely formed thoughts: The recordings were made during a typing task.¹² Brain2Qwerty v2 builds on v1, whose study appeared in Nature the same day.¹ Meta compares its result with roughly 8% word accuracy for earlier noninvasive methods.¹ Testing took place under controlled MEG conditions. Whether the method works as well outside the lab remains unclear.¹³