Top Stories
Claude Sonnet 5 Lands as Anthropic’s New Mid-Tier Agent Model
Anthropic released Claude Sonnet 5 on Tuesday, June 30, 2026, positioning it as a lower-cost alternative to Opus 4.8 for running AI agents. Per TechCrunch, Sonnet 5 launches at $2 per million input tokens and $10 per million output tokens through August 31, after which rates rise to $3 and $15. Anthropic pitches it as cheaper than Opus 4.8, GPT-5.5, and Gemini 3.1 Pro, and as the new default on its free and Pro plans.
The “cheaper” framing is complicated by a tokenizer change. Simon Willison’s developer breakdown finds the new tokenizer produces roughly 30% more tokens than Sonnet 4.6 for the same English input - an effective price hike of about 28-42% depending on language, even though the published rates are unchanged. Sonnet 5 also drops temperature, top_p, and top_k sampling parameters and ships with a 1,000,000-token context window and 128,000-token maximum output. Anthropic told TechCrunch the model is less capable than Opus 4.8 on dangerous cybersecurity tasks and that “Opus 4.8 is still the model of choice for higher accuracy” on agentic work.
US Drops Export Controls on Anthropic’s Mythos and Fable
The US Commerce Department, via Secretary Howard Lutnick, lifted its June 12 requirement that Anthropic obtain an export license for the Mythos and Fable models, according to TechCrunch. Under the deal, Anthropic will restore public access on Wednesday, July 1, and agreed to “proactively detect and address security risks,” collaborate with the government on release protocols, and report malicious activity. Anthropic had already pledged to do much of this voluntarily months earlier.
The reversal ends a roughly two-month restriction: Mythos had been limited to selected organizations since April over vulnerability-exploitation concerns, and Fable launched in June under extra guardrails. Pressure grew as Asian startups released comparable models such as Fugu and Tulongfeng during the ban. Cybersecurity experts quoted in the piece were skeptical the restrictions were genuinely security-driven, reading them as political bargaining chips instead.
Claude Science Targets Drug Discovery and Lab Workflows
Anthropic also unveiled Claude Science on the same day, a standalone product that extends the Claude Code pattern into scientific research. It writes and runs code on compute clusters, prioritizes reproducibility so scientists can trace figures back to their sources, and interfaces with tools used in genetics, chemistry, and protein biology.
The launch event featured pharmaceutical executives, biotech founders, and academic researchers. Anthropic demonstrated Claude Science autonomously identifying new drug candidates for phenylketonuria. The product is available to all paid Claude subscribers and sits alongside Claude Code and Claude Cowork as a flagship product, replacing earlier life-sciences plug-ins.
Proton’s Lumo Chatbot Gets Image Tools and Memory in 2.0
Proton has shipped Lumo 2.0, adding image recognition and generation, persistent memory for user-controlled “Projects,” a “thinking mode” for complex problems, and a claimed 76% speedup. Lumo keeps its zero-access encryption model: data is encrypted in transit and at rest, only the user holds the keys, no server-side session logs are retained, and customer data is not used to train AI models or shared with third parties.
The privacy framing is the differentiator against Gemini and ChatGPT. Proton says Lumo 2.0 is now “roughly equivalent” to mainstream chatbots on raw usefulness, with a free public tier and Plus and Professional paid tiers. For readers waiting for a credible privacy-first assistant that does not sandbag features, Lumo 2.0 is the closest current option on the market.
Henrico County’s 37 Data Centers Push Schools to Cut Power
Henrico County, Virginia has 37 data centers and is asking schools and county workers to conserve electricity in response to a 25% rate hike that will add roughly $5 million to its power bill next fiscal year. The internal email asked workers to close the blinds and turn off their computers.
It is a concrete local example of what AI compute growth looks like on a public balance sheet. The same dynamic is playing out in other Loudoun- and Prince-William-style data-center corridors, but Henrico’s email is the rare instance where a public agency has been forced to put the trade-off in writing rather than absorb it.
Google Ships Nano Banana 2 Lite at Sub-Cent Pricing
Google has released Nano Banana 2 Lite, a faster, cheaper image generator aimed at high-volume workflows. The model produces images in about four seconds at $0.034 per 1,000 images via Google AI Studio, the Gemini API, and the Gemini Enterprise Agent Platform. Google says Nano Banana 2 Lite is the replacement for the original Nano Banana, which is now labeled its “legacy model.”
The wider release also includes Gemini Omni Flash, priced at $0.10 per second of video output, and a new demo app, Omni Product Studio, for turning static product images into e-commerce videos. Nano Banana Pro remains Google’s top-tier option for image quality.
Etched Hits $5B Valuation With $1B in Inference Chip Orders
Etched, an AI chip startup founded in 2022 by ex-Harvard Thiel fellows Gavin Uberti and Robert Wachen, has raised $500M at a $5B post-money valuation and reports $1B in booked orders for its purpose-built inference systems. The TSMC-made chips ship as full “frontier inference clusters” - silicon plus custom racks and software - and Etched claims they beat general-purpose GPUs from Nvidia on speed, cost, and power efficiency for inference workloads.
The round was led by Stripes, with VentureTech Alliance, Jane Street, Hudson River Trading, Two Sigma, and Ribbit Capital participating, bringing total funding to $800M. Angel investors include Andrej Karpathy, Geoffrey Hinton, Fei-Fei Li, Arthur Mensch, Scott Wu, Stanley Druckenmiller, and Peter Thiel. A real revenue line for an Nvidia challenger is rare; the next data point is whether those orders ship and convert.
”Caveman” Plugin Strips AI Output to Cut Token Spend
A new open-source “caveman” plugin rewrites verbose LLM output into terse, plain-speech responses (“Hulk smash” instead of long apologies) for Claude Code, Codex, and Gemini. A senior OpenAI employee contributed code adding Codex support, and developers at OpenAI, Nvidia, and GitHub are reportedly using it.
The plugin is a symptom rather than a cure: teams are reaching for stylistic workarounds because raw inference spend remains volatile. 404 Media cites Accenture finding that PDF-to-slide conversion drives much of the spend. The tool itself has no published benchmarks, and the article does not quantify savings, so treat the cost claims as adoption signal only.
Quick Hits
-
OpenClaw on mobile: The open-source AI agent is now on iOS and Android, paired with an OpenClaw Gateway routing layer; the project has had security and impersonation issues since its earlier viral run.
-
X launches MCP server: X has shipped a hosted MCP server so assistants like Claude, Cursor, and Grok Build can authenticate with a user’s account and search posts, read users, and analyze trends. Write API is not supported.
-
OKX opens an agent marketplace: AI agents can now find jobs, hire each other, and settle payments in stablecoins via OKX’s Onchain OS, with partner GenLayer providing dispute resolution. A closed beta ran with 50 providers.
-
Amazon’s $1B FDE org: AWS has committed $1B in internal resources to a forward-deployed engineer team focused on deploying AI agents inside customer companies, following OpenAI’s $4B and Anthropic’s $1.5B FDE plays.
-
ScarfBench published: IBM Research has released ScarfBench, a benchmark for AI agents on Spring, Jakarta EE, and Quarkus migrations; even the strongest frontier coding agents managed less than 10% behavioral success on whole-application migrations.
-
Wayve tender offer: The UK self-driving AI startup is running an $85M employee tender at an $8.5B valuation, set during its February Series D, ahead of robotaxi pilots with Uber later this year.
-
“AI coworker” framing hurts oversight: Boston University research covered in MIT Tech Review’s Download finds managers caught 18% fewer errors when AI tools were framed as agentic coworkers rather than chatbots, adding weight to oversight-reduction concerns.
Worth Watching
The real price of Sonnet 5. The published rate matches Sonnet 4.6, but a tokenizer change pushes effective cost up 28-42% depending on language. Expect Simon Willison-style developer accounting to collide with Anthropic’s “cheaper agents” marketing over the next billing cycle.
Lumo vs mainstream assistants. Proton has matched the basic feature set - image tools, persistent memory, a thinking mode - on top of its zero-access encryption model. The next data point is whether mainstream assistants start matching Proton’s privacy posture, or whether Lumo holds the privacy-first niche alone.
Data-center power bills. Henrico County’s email is a single county, but the math scales. Watch for similar messages from school districts and municipal agencies in other data-center-heavy clusters, and for state-level regulatory pushback.
Mythos public release. With export controls lifted, Anthropic restores public access to Mythos on Wednesday, July 1. The release is the test of whether the prior two-month restriction made the model more or less competitive with Asian peers that shipped during the ban.