Anthropic just made Claude cheaper

Hi there, this is your daily ☕️ Devshot.

Anthropic cuts Claude Fable price 20% LINK

• Anthropic released Claude Fable 5.2 and Mythos 5.2, faster versions of its most powerful models, and cut Fable's pricing 20% to $8 per million input tokens and $40 per million output tokens, down from 5.1's $10 and $50.
• Both models answer up to 30% faster than the 5.1 versions, with an optional fast mode up to 2.5 times quicker at $16 and $80; Mythos 5.2 is the same model with looser cybersecurity and biology safeguards, limited to vetted organisations.
• Fable 5.2 is open to all users from today on Anthropic's platform plus AWS, Google Cloud and Azure, and blocks about 55% fewer harmless requests than 5.1, though Anthropic still recommends starting with Opus 5.5 for most coding and knowledge work.

DeepMind AI verifies Signal's encryption LINK

• Google DeepMind used an AI agentic framework called AlphaProof Code to mathematically prove key security properties of the actual code behind the Signal protocol, the encryption protecting billions of conversations in apps like Signal and Google Messages.
• Human experts translated Signal's security goals into formal specs, then Gemini Carbon-powered agents generated over six million lines of Lean to prove them, flagging just one percent of the codebase for human review while Lean checked the rest.
• The proofs covered compiled code running on devices rather than only the design, finished in four weeks instead of the usual years, and correctly flagged two previously fixed bugs as insecure when they were reintroduced into an isolated build.

Meta launches Muse Ember AI model LINK

• Meta released Muse Ember, its smallest and most efficient model aimed at the repetitive work agents do all day, tool calls, coding subtasks, routing, extraction, and long document parsing, competing with Claude Haiku and GPT Luna.
• For prompts under 200k tokens it matches rivals at $0.10/million input and $0.50/million output, but pricing climbs above that threshold; on Artificial Analysis's intelligence-vs-cost chart it sits on the efficiency frontier, with no model scoring higher per task.
• Ember offers six reasoning levels from minimal to max, though the author found reasoning level didn't cleanly map to output quality, with results comparable to Muse Wonder at roughly 40x the price.

Effect 4.0 ships 5x smaller bundles LINK

• Effect 4.0, the TypeScript framework for typed errors and structured concurrency, shipped as the first stable release of the line, rewriting its fiber runtime from scratch for smaller bundles, lower memory, and higher throughput.
• Per the team's own benchmarks, a minimal program using Effect, Stream and Schema dropped to 7.1 kB from 35.6 kB, a 5x cut, while 50,000-fiber heap use fell ~86% and throughput rose roughly 6.4x, though the figures remain unreproduced independently.
• The core now ships with zero runtime dependencies and a single shared version, and a new LTS policy runs bug fixes until September 2029; teams on 3.x should start with the v3-to-v4 migration guide, which coding agents can largely automate.

Tool runs 284B DeepSeek-V4 on CPU alone LINK

• A new set of kernel patches, quantization transcoders, and runtime tools can run the sparse 284B-parameter DeepSeek-V4 MoE checkpoint entirely in system RAM on x86_64 and ARM64 CPUs, no GPU required.
• It compresses the checkpoint enough to fit the full model plus KV cache into the ~62.5 GiB usable on a 64 GB machine with zero swap, evaluating all 6 routed experts plus 1 shared expert per layer without pruning.
• Custom cache-aware SIMD tiling, register-stationary FlashAttention, and self-speculative drafting keep decode fast, and you can optionally drop experts to trade a small perplexity hit for higher throughput; submit CPU scorecards to the leaderboard.

Tutorials & Advice

> Modern CSS Might Make Your Front-End More Complex: Replacing JavaScript with modern CSS can shrink one file while making the whole app harder to understand, so verify a native feature truly replaces the old code without disturbing surrounding layout and styling rules.

> TDD With AI Agents: Design the Tests First, Let the Agent Implement: Write the tests that define what a feature should do first, then let the AI agent write code to pass them, so the tests check your intent rather than the agent's own work.

> My Paranoid Coding Machine Setup: Isolate autonomous AI coding agents inside a locked-down virtual machine, letting them run freely while limiting damage from accidents or supply-chain malware.

> Teach AI coding assistants your design system with Markdown: Document your component rules and design tokens in Markdown beside the code so AI coding assistants follow your conventions instead of inventing inconsistent UI.

> How to add authentication to a Rust axum app with WorkOS AuthKit: Explains how to add login to a Rust web service by using WorkOS AuthKit to seal session cookies, verify tokens locally, and authenticate webhooks.

Other news you might like

• Extending LLDB to Debug Applications That Span CPUs and GPUs LINK

• Pinterest and Intel Take Open-Source Profiling to Internet Scale LINK

• JIT bitmap text rendering for microcontrollers LINK

• Mesa 26.3 Released With VK_AMD_anti_lag For RADV, Many Other Vulkan Improvements LINK

• Locating the source of a mysterious beep with Claude LINK

• Open-source GPU task manager reveals exactly what's eating your VRAM LINK

• Broadcom open-sources its Kubernetes VM manager LINK

Trending tools

Die With Me: a Mac app that tracks Claude and Codex usage, compares it with friends, and opens a shared chatroom during AI cooldown periods. LINK

Jev: adds structured AI-driven decision-making to automation workflows, letting developers build reliable branching logic without hand-coding every rule LINK

beacon.​sh: open-source persistent memory for AI coding agents, letting them retain context across sessions so they learn from past work instead of restarting. LINK

Scalekit: connects AI agents to 30,000+ tools across 600+ apps through one MCP gateway, handling auth, scopes, and per-user permissions for production use. LINK

OpenSwarm: an AI-first operating system where agent swarms run the machine, letting you orchestrate dozens of parallel tasks instead of working sequentially LINK

Web Search Agents by Nimble: domain-specific crawling agents that self-learn your use case to pull deeper research for company enrichment, regulatory tracking, and ongoing monitoring tasks. LINK

Trending papers & reports

Android phone forensics show that onscreen dates can predate a device's real setup by a median of ~3 months, so investigators risk misdating evidence unless they pin when a handset's own life actually began. LINK

Testing for AI agent teams reveals that standard test suites miss many coordination faults, catching under 60% of unsafe tool calls and spotting removed safety restrictions only ~20% of the time. LINK

Executable specifications are proposed as checkable rulebooks that independently verify quantum and machine-learning software behaves as intended, with the paper mapping out the key hurdles blocking this approach from working today. LINK

Browser voice synthesis delivers studio-quality speech in twelve languages including Sanskrit entirely on-device, starting audio in under a second instead of several, using compressed voice files small enough for phone and desktop browsers. LINK

AI-written CAD designs can look finished yet silently flunk real requirements ~14% of the time, but forcing the agent to measure before claiming done cut silent failures to ~3% and lifted trustworthy claims from 64% to 96%. LINK

See you tomorrow for a new dose of ☕️ Devshot!