AI News
| General
The AI model race is picking up speed: Anthropic has launched Sonnet 5.5 and Haiku 5.5, Meta has unveiled Muse Wonder and Wonder Full, and Google has introduced Gemini 4 Argon, giving users and developers a fresh lineup to compare.
By Klara Pelhe
Oct 15th, 2026
klarapelhe.rs/en/ai-news
Anthropic, Google and Meta Unveil New Models as the AI Race Heats Up
Anthropic released Claude Sonnet 5.5 on 28 September for coding and professional document work. It targets bounded tasks, while Anthropic favors Opus 5.5 for open-ended judgment.¹
Anthropic’s Terminal-Bench 4.0 result rose from Sonnet 5’s 10.3% to 70.6%. The test measures agents completing command-line tasks; this vendor result is not a universal success rate.¹
Sonnet accepts text and images within a million-token context—the material held for a request.⁵ Sensitive cybersecurity requests may fall back to Sonnet 5.¹
Sonnet is offered in Claude apps, through an API—a software interface—and on AWS, Google Cloud and Azure.¹⁵ Pricing is $2 input and $10 output per million tokens, the units used to measure model input and output.¹
Independent evaluator Artificial Analysis measured a score of 56 on its Intelligence Index, a composite test of AI capabilities, at maximum effort, alongside roughly 193,000 output tokens and $7.60 per task—about 50% above Sonnet 5’s task cost. Those September tests used launch pricing and a prerelease deployment with a structured-output bug subsequently fixed for release. They establish a costly maximum-effort configuration, rather than today’s bill for a typical coding session.⁵
Anthropic’s new Claude generation is arriving in stages. Opus 5.5 opened the family on 22 September, with Sonnet and Haiku promised next.¹⁶
Simon Willison’s small drawing experiment makes the tradeoff tangible. Asked for a pelican riding a bicycle as an SVG—a vector image described in code—Sonnet exhausted 128,000 reasoning tokens at maximum effort without producing the image, costing $1.28. At xhigh it returned an image in 41 seconds for 5.74 cents. More thinking was actively unhelpful in that example.¹⁷
Agents use tools to carry out tasks, and the effort setting controls how much the model reasons along the way. Anthropic’s migration guide recommends medium for clearly specified work and high for harder tasks; the API defaults to high. It also changes an integration detail: disabling reasoning before a response now uses between_tools, accepted only through high effort, rather than the old disabled setting. Moving an application requires checking behavior and costs as well as replacing its model name.¹⁵
Haiku 5.5 arrived on 7 October as the first Haiku with adjustable reasoning effort, targeting high-volume summaries, classification and compaction—the shortening of an agent’s accumulated conversation. It can act as a subagent, handling a bounded piece of work for a larger model.² It is offered through Anthropic’s platform, AWS, Google Cloud and Azure as claude-haiku-5-5.¹⁸
Haiku’s million-token context has a pricing threshold at 100,000 prompt tokens. These rates are dollars per million tokens; longer prompts enter the higher tier. Cache reads reuse previously processed input.⁶
Anthropic has returned to its small-model family almost a year after Haiku 4.5, which launched on 15 October 2025. That earlier model was built for fast responses and lower-cost work, including tasks delegated by a larger Claude model.¹⁹
The tiered rates are also documented in Willison’s release analysis. He found that his same long prompt used about 1.25 times as many tokens with Haiku 5.5 as with Haiku 4.5, illustrating how a tokenizer—the system that splits content into billable units—can change the comparison. A nominal tenfold price reduction on short prompts is not automatically a tenfold reduction for identical work.²⁰
Artificial Analysis measured an Intelligence Index score of 43 at maximum effort and around 162,000 output tokens per task. Its provisional costs excluded the long-prompt price step.⁶
A hypothetical document workflow could give Haiku individual extraction or summary tasks, then let Sonnet reconcile contradictions across the results. That keeps the delegated work bounded rather than assuming the smaller model can independently complete every complex project.
The same launch halved Sonnet 5.5 cache reads from $0.20 to $0.10 per million tokens. Anthropic estimates roughly 20% savings on most agentic work, with the actual benefit depending on how much input is reused.²
Meta launched Muse Wonder and Muse Wonder Full on 12 October. Active Memory retrieves relevant information across conversations. Active Focus describes continued work through obstacles, with clarification and approval for consequential actions. These are Meta’s descriptions of intended behavior: a planning assistant, for example, could reuse preferences rather than need the same briefing repeatedly.³
Developers can use Wonder through Meta Model API and Muse Code, Meta’s terminal coding tool. Its API lists $4 input, $0.20 cached input and $20 output per million tokens. Wonder Full is offered through Muse Code subscriptions; the model page lists it in the $50-per-month Power Usage plan, with a $200 Maximum Usage tier above that.²¹ The consumer Muse rollout was announced as forthcoming, starting with subscribers.³
The model card specifies text, image, video and audio input, text output and a million-token inference context. Meta’s own comparisons generally use maximum reasoning effort, and some reported external results retain different providers’ agent setups. The card therefore supports a description of tested capabilities, without establishing that every remembered preference or long-running task will be handled correctly.²²
Meta introduced its Muse family with Spark on 8 April, a model that already combined multimodal reasoning with tool use. Wonder follows roughly six months after that debut.²³
Google introduced Gemini 4 Argon on 30 September, its new flagship reasoning model for complex software engineering and sustained, multistep work. The model is designed to tackle demanding tasks that require planning and reasoning across a long sequence of steps.⁴²
Argon accepts text, images, video, audio and PDFs, and returns text. Its 1,048,576-token window is shared between input and output. Ordinary synchronous responses stop at 65,536 output tokens; longer generation uses a separate long-decode mechanism. Image and audio generation are not included in this model.²⁴
Google initially restricted access to trusted cybersecurity defenders in its Fairwind program.⁴
Google opened Gemini 4 Argon on 13 October to developers through its paid API, to Pro and Ultra subscribers in Gemini and AI Mode in Search, and to Ultra subscribers in Antigravity, its coding-agent environment. “Available for everyone” describes these broader access routes, not a free model.⁷
Introductory API pricing is $2 input and $10 output per million tokens through 30 November; afterward, Google lists $4 and $20.²⁵
Google describes an 11-hour performance pass over Agent Development Kit, its open-source framework for agents. Engineers supplied repository rules, shipped two proposed optimizations and rejected four, including a potentially stale cache. The accepted changes reduced in-memory session preparation overhead by about 40% in long sessions and package import time by 21%; time from a fresh process to its first model request stayed effectively flat.²⁶
The merged commit changes session copying and lazy package loading, tying the demonstration to inspectable code. Those improvements concern framework overhead in the tested setup, rather than a 40% speedup for every AI application or the model itself.²⁷
Developer Theo Browne reported a sharply different experience in his own codebases on 14 October: Argon repeatedly invented details about the systems it was modifying, making it unusable for his work. His thread supplies a practical failure mode that the successful Google example does not erase.⁸ Google’s account, in turn, describes human review and repository-specific instructions, conditions that matter when comparing the experiences.²⁶
For long decode, Google’s API continues a response in hops of up to 64,000 tokens. Background execution manages those continuations on the server; its output budget includes thinking as well as the final answer. A million-token window is consequently a capacity limit, rather than a promise of a million-token answer delivered in one ordinary request.²⁵
OpenAI began rolling out Intelligent UI in ChatGPT on 7 October, allowing answers to contain interactive tools. A written explanation can tell you how to split a dinner bill; a calculator lets you change the numbers and see the result. Responses can contain tools people use directly in the conversation, including a bill-splitting calculator. Behind the interface, a native component library and a compiler display each part progressively as the model generates it.⁹
The rollout began on paid Chat tiers and expanded to Free and Go on 8 October; workplace access depends on administrator settings. OpenAI’s release notes distinguish the Chat models—Sol on paid tiers and Luna on Free/Go—from models used in Work and Codex, OpenAI’s task and coding agents, which this change did not replace. The separately selected Pro reasoning model, Astra, does not support Intelligent UI.²⁸
OpenAI’s Thibault Sottiaux announced on 12 October that GPT-6.1 Sol’s default context in Codex and Work increased from 272,000 to 500,000 tokens, with a configurable maximum of one million. More room can postpone compaction, but it can also mean processing more input. This is an operating-window change for an existing model, rather than a new model launch or a guarantee of better results.²⁹
In a long coding session, earlier requirements and tool results compete for space with new code. A larger context lets an agent retain more material before compressing the conversation.²⁹
On 14 October, Anthropic’s Thariq Shihipar highlighted a desktop update to its coding agent Claude Code: faster session starts, sessions visible side by side and easier splitting of work. When several coding tasks run at once, following each session becomes part of the work. His post also describes Claude initiating subsessions when useful. The concrete benefit is coordinating separate pieces of a coding task; the announcement does not establish how reliably those pieces are reconciled, and its informal “v2” label should not be read as a formal product version.³⁰
Google released EmbeddingGemma 2 on 6 October under Apache 2.0. It turns text, code, images, audio and video into embeddings: numerical vectors that allow content to be searched by meaning. The full model has 740 million parameters, learned numerical values; text-only operation uses 270 million. In Google’s Pixel 11 Pro setup with quantized, reduced-precision weights, text weights required about 191 MB of active memory and full multimodal weights about 567 MB, excluding the rest of an application.¹⁰
Imagine a research folder containing notes, recordings and video. A local tool could use a spoken question to retrieve related notes and video excerpts, then give those passages to a generative model. That is retrieval-augmented generation, or RAG: supplying retrieved material as evidence for an answer. Keeping embeddings local does not keep a subsequent cloud generation step local.
The model card explains an engineering tradeoff: 768-dimensional vectors can be shortened to 512, 256 or 128 dimensions to reduce storage. Queries and documents must use matching dimensions, and truncated vectors need normalization. Google warns that 128 dimensions substantially degrades multimodal quality. Its shared 8,192-token input budget also means text and media compete for space; the advertised audio, image and frame maximums cannot all be combined in one input.³¹
Microsoft introduced Topics Portfolio in Bing Webmaster Tools on 13 October, turning retrieval queries into a broader map of a site’s coverage. Bing already showed publishers which pages its AI answers cited. The beta groups retrieval queries by topic and highlights questions citing other sites but not yours, suggesting gaps a publisher could investigate. Totals are approximate and Bing-specific; grounding queries are not users’ original prompts. AI classification can misgroup topics, and insufficient data hides topics or the feature itself. Microsoft calls the figures directional: an opportunity is not a promise of better rankings.³²
OpenAI announced a visual ad test on 5 October, scheduled for later in October in the U.S. with an initial advertiser group. Ads during image generation on Free and Go would be labeled and kept separate from the generated image. OpenAI says ads do not influence answers. This is paid placement, a different mechanism from a publisher earning an organic source citation. The announcement does not establish availability for Serbian advertisers or users.³³
For advertisers, appearing beside an AI conversation raises a different question from earning a source citation: how to buy and measure that placement.³³
Its expanded measurement tools include a tracking pixel and Conversions API, which connect events such as purchases or sign-ups to campaigns, plus attribution partners. Advertisers can include a one-day view-through reporting window for conversions after an impression without a click. That assignment is not the same as demonstrating an additional purchase caused by an ad; OpenAI says its experiments across geographic areas to estimate added purchases are still early. Its selected partner case studies therefore do not establish a general return on advertising spend.³⁴
On 6 October, OpenAI published mathematical manuscripts produced by an unreleased internal model, together with supporting artifacts and ten reasoning summaries. Many proofs have Lean formalizations: machine-checkable statements and proofs written in a proof-assistant language. The company says the average result consumed compute equivalent to roughly three hours of ChatGPT Pro thinking. That is a compute comparison, not a statement that the same model is available to Pro users.³⁵
A model’s claim to have proved something becomes useful to mathematicians when they can follow and check the argument. OpenAI’s latest release puts more of that material in public.³⁵
The repository explicitly distinguishes stages of verification and warns that unformalized results may contain issues. It preserves revisions and gives manuscript-specific citation instructions. A computer-checked formal proof can strengthen confidence in the formal statement, while mathematical significance, novelty and the relationship between that statement and a paper’s broader claims still require scrutiny. Publication supplies material to examine, rather than validating every announced result at once.³⁶
Z.ai introduced GLM-Legal on 12 October, specializing GLM-5.3-Flash through further legal training and reinforcement learning, which rewards desired task outcomes, in tens of thousands of legal task environments. It also iterates on the agent harness—the tools, context and feedback around the model—with practitioner input. The aim is producing legal work from documents and instructions, rather than merely answering legal trivia. Its announced AutoClaw desktop integration was still forthcoming.³⁷
A legal team working from a bundle of documents needs a usable draft that follows instructions and stays tied to its sources. GLM-Legal is aimed at that kind of work.³⁷
Vals AI’s dated evaluation update reports 19.17% on Harvey’s Legal Agent Benchmark and lists API pricing of $0.075 input and $0.25 output per million tokens.³⁸ That benchmark gives agents files, editing and shell tools and document-production skills, then assesses task-specific requirements on a held-out set. Passing individual criteria differs from fully passing a task; a roughly 19% strict all-pass result is not “19% accurate legal advice,” nor evidence of dependable unsupervised legal practice.³⁹
AI evaluator METR proposed on 14 October that outside assessors examine internal agent behavior, monitoring and AI-driven research acceleration, using internal transcripts, systems and private staff interviews. The question is whether a lab would detect agents acting against instructions. METR says immature science and absent agreed safety standards prevent confident safety assurance.¹⁴
An AI lab’s safety also depends on agents working behind the scenes. METR wants outside researchers to examine that internal work.¹⁴
Its companion policy identifies a structural limit: companies voluntarily choose access and can redact information or end an engagement. METR seeks the right to disclose those restrictions and publish its own findings, but acknowledges that independence remains difficult. A report’s access conditions and omissions are therefore part of its substance, rather than administrative details.⁴⁰
U.S. prosecutors announced charges on 13 October against Abdulrahman Alfozan, a 35-year-old Saudi national who had worked as an OpenAI engineer. He was arrested on 9 October. The allegation concerns a concealed recording device installed behind a television around 28 September and discovered on 30 September in an executive conference room. Reuters reports that OpenAI terminated his employment and referred the matter to the FBI. These dates separate the alleged conduct from its later public disclosure.⁴¹
AI security also involves protecting private conversations. An allegation involving OpenAI has brought that concern into a criminal case.⁴¹
According to the complaint summarized by the Justice Department, the room hosted discussions of confidential research, product plans and government relationships. The complaint also alleges searches about meetings and chatbot queries about constructing and concealing a recorder. These are prosecution allegations; Alfozan is presumed innocent unless proved guilty.¹¹
The political framing is explicit. U.S. prosecutors and the FBI describe protecting American AI innovation as an economic and national-security priority.¹¹ Officials are presenting an alleged workplace interception as part of the strategic competition over AI capabilities. That framing does not establish a foreign-government operation.
AFP reports that the Justice Department did not accuse Saudi Arabia of spying in its statement. What, if anything, the device recorded was still unclear. OpenAI said it had no indication of inappropriate access to or exfiltration of model weights, source code or customer data—a company assessment, not an independent forensic conclusion. Nationality alone supplies no evidence of Saudi state involvement; motive, recipients and the extent of any information exposure remain unresolved.¹²
OpenAI announced optional, default-off API watermarking for selected models globally on 5 October; eligible EU ChatGPT and Codex text would follow over several weeks. Detector access initially requires researcher or expert-organization approval. Completion was not announced.¹³
When an AI answer is copied elsewhere, readers may lose the clues to its origin.¹³
The textGrain watermark is a statistical pattern in token choices, not a visible label or hidden character inserted into a document. The technical report describes generation influenced by secret-key randomness, with a budget limiting how much randomness is removed. A detector reconstructs the keyed signal from the text; it is looking for evidence of that watermark rather than determining the truth of the content.⁴²
In OpenAI’s 400-token English test, at a 1% target false-positive rate, replacing 10% of words with synonyms reduced detection from about 92% to 66%; replacing 25% reduced it to 17%.¹³
An absent watermark cannot establish human authorship. Treating a positive signal as proof of ownership or a complete editing history would go beyond what this statistical mechanism measures.⁴²