Sonnet 5 ships, Zeta cuts prediction tokens 67% — Dev Signal
July 7, 2026
Sonnet 5 ships, Zeta cuts prediction tokens 67%
Share:
Tool of the Week
Claude Sonnet 5 launches with 30% higher token costs
Sonnet 5 matches Opus 4.8 performance at lower list prices, but new tokenizer inflates actual costs by ~30% for English text; adaptive thinking enabled by default changes inference behavior.
If you're running Sonnet 4.6 in production, switching to Sonnet 5 requires tokenizer recalculation and explicit thinking configuration—apparent price parity masks real cost increase. The 1M context window and 128k output tokens unchanged from prior versions.
Sonnet 5 replaces Sonnet 4.6 as the mid-tier option, but the tokenizer change means token budgets and cost projections break. Requires testing with your actual workloads using the new token counter before migration. Skip unless you need the performance bump—otherwise stay on 4.6 through August discount period to reset baselines.
“its performance is close to that of Opus 4.8, but at lower prices”
“The same input text produces approximately 30% more tokens than on Claude Sonnet 4.6”
“Adaptive thinking is on by default, unless you specify "thinking": {type: "disabled"}”
claudepricingtokenizationapi-changescost-analysis
Dev Signal
Get issues like this in your inbox — free, every weekday.
Multi-Region prompt format reduces output tokens from ~270 to ~90, dropping response latency 28% (p50) and server overhead by 30%.
Faster edit predictions on keystroke directly improve interaction latency in Zed; lower token emission means cheaper self-hosted deployments and reduced inference costs at scale.
Replaces Zeta2 as default in Zed today. Open-weight model available on Hugging Face with Rust PyPI bindings for local inference. Ready now: download and run locally or use Zed Pro/Business tiers. No migration friction—it's the default.
“Zeta2.1 emits 3x fewer output tokens than Zeta2, bringing predictions up to 50ms faster”
“Output tokens (avg)~270~90 (−67%)”
“Response Time (p50)189ms136ms (−28%)”
“requiring 30% fewer servers to serve the same traffic”
“Zeta2.1 is open-weight, just like Zeta1 and Zeta2”
“Zeta2.1 is the default edit prediction model in Zed today”
Peewee 4.x adds native asyncio support via greenlets to execute_sql, unified cross-backend JSONField, and declarative relation eager-loading—enabling true async/await with FastAPI without threadpool workarounds.
Eliminates forced sync-to-async bridges (threadpools, sync_to_async serialization) that block concurrent request handling. Developers can now write fully async FastAPI apps with ORM queries that actually await cleanly on the event loop.
New benchmark evaluates text-to-image models on what actually matters for scientific figures: label legibility, semantic correctness, and disciplinary conventions—not photorealism.
Current image-generation benchmarks measure natural-image quality (compositionality, object counting, photorealism) but miss critical failure modes in scientific diagrams: garbled labels, incorrect entity relationships, and convention violations. SciDraw-Bench's four-dimensional evaluation (text fidelity, semantic correctness, structural quality, convention adherence) gives you concrete metrics to assess whether a model can replace manual figure creation in your domain.
Don't adopt yet for production. SciDraw-Bench is a research paper introducing evaluation methodology and a domain-specific baseline (SciDraw AI); no public model release or API is mentioned. Actionable only if you're training custom figure-generation models and need a structured evaluation framework. The benchmark itself is reproducible; implementation requires access to their task specifications and evaluation code (pending release).
“None of them measure what makes a generated scientific figure usable: correct and legible text labels, faithful depiction of entities and their relations, coherent diagrammatic structure, and adherence to disciplinary drawing conventions.”
“text fidelity remains the hardest dimension for all systems”
“the domain-specific system substantially outperforms the general-purpose baselines on every dimension and figure type”
Replaces backend-specific playhouse JSON extensions and sync-only query patterns. Requires Postgres 9.2+/MySQL 8/SQLite 3.38+ and explicit AsyncPostgresqlDatabase/AsyncMysqlDatabase imports. Ready now—stable release with full FastAPI/Starlette/Quart integration patterns documented.
“Peewee takes the query execution coroutine, switches control to the async layer and awaits it on the event loop”
“the core JSONField provides a consistent API that works across SQLite (3.38+), Postgres (jsonb), MySQL 8 and MariaDB”
“Connections in pwasyncio are task-local and maintain their own transactional state, so transactions are isolated and won't interleave across concurrently-running tasks”
“Peewee 4 is broadly compatible with the 3.x releases”
peeweeasyncioormjson-fieldfastapi
Claude Code Chrome extension automates visual testing
Claude can now control your browser, take screenshots, and iterate on UI changes autonomously—removing manual test-and-verify cycles from feature development.
Developers can offload repetitive visual validation and layout refinement to Claude, compressing the build-test-debug loop from hours to minutes. This shifts focus from mechanical checking to architectural decisions.
Replaces manual browser testing and screenshot validation for UI work. Requires Node.js 18+, standard Chrome (not Dev), and Claude Code installed. Worth trying now if you're already using Claude Code—low friction to add to existing workflows, but effectiveness depends on how precisely you specify measurable requirements.
“Claude builds features, tests them visually in Chrome, spots issues, and iterates without you manually checking each change”
“If you can't verify it by looking, Claude can't either. Try to specify measurable requirements”
“Claude Code won't detect the extension in Chrome Dev even when installed”
CVE-2025-23166 allows attackers to crash Node.js processes via malformed crypto inputs; upgrade immediately to 20.19.2+, 22.15.1+, 23.11.1+, or 24.0.2+.
Untrusted cryptographic operations are standard in production apps; a single malicious input can take down your entire runtime. HTTP/1 request smuggling (20.x only) also bypasses proxy-based access controls, exposing backend services.
Replace your current Node.js version with the patched release for your line (20.x, 22.x, 23.x, or 24.x). No code changes required. Do this today—the high-severity crash vector affects all active release lines in production.
“Improper error handling in async cryptographic operations crashes process (CVE-2025-23166) - (high)”
“Such cryptographic operations are commonly applied to untrusted inputs. Thus, this mechanism potentially allows an adversary to remotely crash a Node.js runtime.”
“This vulnerability affects all users in active release lines: 20.x, 22.x, 23.x, 24.x”
“Releases will be available on, or shortly after, Wednesday, May 14, 2025.”
node-jssecuritycve-2025-23166cryptoremote-crash
Mistral Vibe unifies work and code agents
Single agent now handles multi-step admin tasks, research, and end-to-end coding workflows—from request to merged PR—across web, IDE, and CLI with isolated sandbox execution.
Reduces context switching between separate tools for knowledge work and coding; sandbox isolation and visible tool calls let you inspect diffs before approval, lowering deployment risk in agentic workflows.
Replaces Le Chat + separate coding tools. Requires GitHub/Slack/Google Workspace connectors and VS Code extension install. Code Mode on web is ready now; Slack trigger integration ships June. Start with Free tier for testing, Pro ($14.99/mo) for coding workflows.
“Vibe is now one agent for long-running, multi-step work”
“takes coding work from request to merged change, across the web app, your editor, and your terminal”
“the agent runs in an isolated sandbox, and you manage sensitive actions and inspect diffs as it's writing code”
“Vibe deeply grounds its work in your context, reaching across Google Workspace, Outlook, SharePoint, Slack, GitHub, and any custom connectors”
“Sessions can run in parallel, can persist while your machine is off, and can be triggered from third-party apps, such as Slack (coming in June)”