Signal, not noise. The models, tools, pricing changes and ecosystem shifts that actually matter for developers — curated once a week.
A coding-and-agent model that lifts DeepSWE from 49% to 65.3% and AutomationBench from 17% to 30.4% — at $0.75 / $3.75 per M, half price through 2026. It also becomes the base of Google's personal agent, Gemini Spark.
Zhipu's 753B model claims ~50% coding lift and a CyberGym 84.5% that matches closed frontier labs, after finding 2,436 real-world CVEs. Weights open in two weeks under a free security-audit program.
DeepSeek's first multimodal model drops into existing V4-Flash workflows with a one-line model swap — image input at ≤384 tokens each, approaching Opus 4.8 on multimodal agent tasks. Harness 0.1.1 supports it natively.
Research & community-reported scores · TerminalBench 2.1
Research & community-reported prices
DeepSeek's first multimodal model (284B/13B-active MoE, 1M context) takes image input at ≤384 tokens each. Multimodal agent benchmarks approach Opus 4.8 while text ability stays level with V4-Flash — a one-line model swap to upgrade.
Google's follow-up lands three weeks after the last Flash. DeepSWE jumps 49.0% → 65.3%, AutomationBench 17.0% → 30.4%. $0.75 in / $3.75 out per M until end of 2026, then $1.50/$7.50.
Z.AI ships a Turbo variant of GLM-5.2 on Aug 17, an efficiency-tuned sibling to the 5.3 flagship — a sign Zhipu is iterating its mid-tier cadence in the open-weight race.
87.9 vs Claude Code's 86.7 — the first open model to top the terminal benchmark. 1M context, three reasoning levels, OpenAI Responses format natively, off-peak half-price billing.
Unsloth ships GGUF variants sized for small hardware. The 27B becomes a realistic local deployment for private coding and offline agent work, narrowing the release-to-use gap.
Meta's local coding agent runs on 24GB+ VRAM (Ollama/LM Studio support, Q4_K_M at ~17GB) — an execution-layer option that removes the per-token bill for well-scoped tasks.
Harness 0.1.1 natively routes image input to the new V4-Flash-Vision-Exp, letting agents analyze screenshots, charts and PDF pages inline — multimodal agents on the MIT runtime without extra glue code.
The dsh-plugin tag now covers 700+ public repos — agent teams, cross-session memory, Claude Code skill migration, and mini-IDEs. The "everything is a plugin" bet is attracting real builders.
Anthropic's +50% weekly allowance runs through August 31 across Pro, Max, Team and Enterprise — the fourth extension since May. Budget as though it ends September 1.
An AI teammate that signs into tools, does work, and returns completed results — shifting agents from interaction to delegation. Trust controls (scoped credentials, approval gates) remain the open question.
Cursor's plan matrix now spans Start (India ₹649) to Ultra ($200). Daily agent users typically land at $60-100/mo — the most uneven cost profile of the three big tools.
Teams routing planning (Opus 5 / Sol) to expensive models and execution (Sonnet 5, open-weight, local) to cheap ones consistently beat single-vendor stacks — per-subagent model control makes it practical.
$0.75 / $3.75 per M until December, then $1.50 / $7.50. Google is buying developer adoption on the coding tier — priced to undercut both Sonnet 5 and the open-weight crowd.
Anthropic confirmed the September 1 increase to $3/$15 is cancelled — the intro rate stands. Sonnet 5 is now the fixed mid-tier workhorse between Haiku and Opus 5.
The new time-of-day model (V4-Flash $0.22/$0.66 off-peak, $0.44/$1.32 peak) got a follow-up: from Aug 23, Sat-Sun bill at off-peak rates all day — weekend batch/agent runs get the lowest price without scheduling around peak windows.
Image input bills as text at ≤384 tokens per image (~$0.000085 off-peak, ~11,800 images per dollar). A screenshot-driven agent runs a 300-step task for under $0.03 — multimodal at a near-text cost.
SiliconFlow prices V4-Pro at $1.32 in / $3.96 out with cached input at $0.44, and off-peak (valley) hours run at half price from Aug 17 — a new commercial pattern for cheap model tiers.
OpenAI's budget tier rose from $6 (July) to $8. Codex access now reaches from the free tier to Plus ($20) and Pro ($100+) — credit-metered, built for parallel agent runs.
Claude Code meters weekly allowance (predictable ceiling); Codex meters token credits; Cursor mixes allowance with overage — daily agent users routinely exceed $60-100/mo on Cursor.
Model 2 slightly beats Mythos 5 on CoBench (62.8% vs 50.3%) and is already used heavily for coding and agents. The report lists real incidents: multi-agent drift, chain-of-thought leakage, alignment-faking data.
Zhipu offers free audits for critical open-source projects, arguing frontier security capability should be shared, not a closed privilege. Weights land in two weeks under this program.
Google's assistant now powers Gemini Spark, a personal agent spanning Gmail, calendar and docs — the same agent infrastructure now drives Gemini 3.7 Flash's coding tier.
DeepSeek Harness plugins now migrate Claude Code skills, while Codex imports Cursor-managed skills. The "rewrite everything when you switch" era is ending — skills travel with developers.
Three open releases inside two weeks (Qwen3.8-Max, GLM-5.3, V4-Pro) — frontier-class open models are now a rolling release cadence, each shipping license and hardware caveats.
Gemini 3.7 Flash at half price is the clearest "coding agent as commodity" signal yet: a model that gains 16 points on DeepSWE while undercutting Sonnet 5 and the open-weight API tier. The coding layer is being priced toward zero, and Google is willing to buy adoption with a six-month promo. Budget around the 2027 repricing, but enjoy the window.
GLM-5.3 doing security work that matches closed frontier labs — and finding 2,436 real vulnerabilities in real codebases — reframes open weights from "cheap alternative" to "capable enough for production security work". The Open Shield program is also a smart trust play: audits create good PR, real findings, and a reason for enterprises to take open models seriously.
DeepSeek shipping a vision model at a text-like price is the under-the-radar headline of the week. V4-Flash-Vision-Exp lets any existing Flash workflow add image understanding with a one-line model swap and ~$0.000085 per image — the cheapest multimodal entry into agent work yet. The 800px downscale trap on dense small-text images is the real caveat; plan tiling for invoices and full-page PDFs.
V4-Pro topping TerminalBench 2.1 at 87.9 remains a milestone: for the first time an open model leads the terminal-agent benchmark, and DeepSeek pairs it with an MIT harness that already has 700+ plugins. But the same week it raised V4-Flash prices into time-of-day billing — the "cheapest open API" sticker is gone. The new weekend rule (all-day off-peak from Aug 23) is the play: batch CI, test suites and agent sweeps to Sat-Sun and cache aggressively. Scheduling is becoming part of the bill.