News
Updates, product notes and practical coverage for AI platforms and operators.
Anthropic has launched Claude Haiku 5.5 as its fastest and most capable small model, with pricing designed for high-volume workloads. The model adds adjustable effort, stronger computer use and coding performance, broad cloud availability, lower Sonnet cache-read costs and monthly API credits for Claude Max and Team subscribers.
Google has announced Gemini 4 Argon, a frontier model aimed at long-horizon coding, professional knowledge work and cyber defence. Its 30 September release begins with selected security defenders, not general public access. Google claims a one-million-token output limit and has disclosed introductory API pricing, while saying broader availability depends on feedback and stronger safeguards.
OpenAI has released GPT-6.1 Sol for complex coding and professional work, positioning it below Astra on price while retaining a large context window. The 29 September release also brings beta multi-agent support in Responses. We examine the published capabilities, pricing, regional options and the practical evaluation questions before teams switch production workloads.
Upstage lists Solar Mini 4 as released on 22 September. The compact model has 35 billion total parameters, 3 billion active per token, a 512,000-token context and up to 128,000 output tokens. It supports tools and structured responses. Published list rates are US$0.10 input and US$0.40 output per million tokens, with a temporary half-price offer.
Anthropic has launched Claude Sonnet 5.5 with faster output and lower claimed cost per completed task than Sonnet 5. API token rates remain US$2 input and US$10 output per million. The model is available through Claude and major cloud partners, with new effort settings and migration considerations for developers switching from thinking-off configurations.
SpaceXAI has launched Grok 4.7 for coding and knowledge work, with a larger base model, longer training run and new safeguards. It is available through the Grok API, Grok Build and other supported services. Published token prices start at US$2 input and US$6 output per million; teams should test task quality, long-context charges and tool behaviour before switching.
OpenAI has released GPT-6 Sol and GPT-6 Luna through the Responses and Chat Completions APIs. Sol targets complex coding and agentic work, while Luna is priced for focused, high-volume tasks. Both accept text and images and return text. Developers should compare total task cost, long-context pricing and tool support before replacing existing models.
AWS has detailed Grok 4.6 on Amazon Bedrock, adding the standard Bedrock Runtime endpoint and Converse API alongside the OpenAI-compatible Mantle route. The model has a 500,000-token context window and four reasoning settings. Teams should distinguish the September documentation from the earlier August launch, and check endpoint-specific features, residency, pricing and safeguards.
Anthropic has released Claude Opus 5.5, positioning it for agentic coding and knowledge work. The company says typical workloads cost 40 per cent less than Opus 5 at default settings, with output over 30 per cent faster. Those are vendor results: teams should test quality, usage and safeguards against their own tasks before migrating.
AWS has made Moonshot AI’s Kimi K3 generally available through Amazon Bedrock. The open-weight model combines native vision, a million-token context window and explicit prompt caching, with cross-Region inference across Bedrock regions. This is a new hosting option, not a new Kimi model launch; teams should assess regional routing, cost and workload fit.
Google has introduced Gemini 3.8 Live and Live Extended Thinking for real-time voice applications. The models support visual context, multilingual conversation and tool calls that can run while dialogue continues. The Live API release is distinct from rollout across consumer products; benchmark claims and regional access should be assessed against the actual use case.
Baseten has added Inkling Small to its hosted Model APIs and dedicated-deployment service. The open-weight mixture-of-experts model has 276 billion total parameters, 12 billion active parameters, a one-million-token context window and native text, image and audio inputs.
Featherless has added hosted access to Moonshot AI's open-weight Kimi K3 model through its OpenAI-compatible API. The initial configuration offers vision input, FP8 quantisation and a 32K context window, while the provider works towards larger context options.
Liquid AI has released two open-weight LFM2.5 encoder models for classification, retrieval and token-level language tasks. The compact 230-million and 350-million parameter models target local, on-premises and high-volume deployments, with an 8,192-token context window and public evaluation code.
Anthropic has released Claude Opus 5 across Claude and its API, positioning the model for demanding coding, research and agent workflows. The launch keeps Opus 4.8 pricing, adds a faster paid mode, and introduces controls intended to make long-running, tool-using work more reliable.
xAI has released Grok 4.5 for coding, agentic tasks and knowledge work, with immediate access through its API, Grok Build and Cursor. The company is positioning the model around engineering performance, faster generation and lower token use, while its benchmark and efficiency claims still require independent testing.