BriefTechNews
Gemini 4 Neon lands at $3/M input tokens; Carbon's coding slowness dissected

Gemini 4 Neon lands at $3/M input tokens; Carbon's coding slowness dissected

8 min read · 17 sources

TL;DR
  • Google launched Gemini 4 Neon at $3 per million input tokens with a 1M-token context, scoring 61 on the Artificial Analysis Intelligence Index.
  • Gemini 4 Carbon took 22m21s on a coding task where Opus 5.5 took 9m33s, running 11x the actions and 16x the input.
  • OpenAI's models cut circuit area on its Jalapeño chip by 10-13% and delivered up to 14% improvements on AMD's flagship GPUs.
  • Anthropic's compute spending grows 4x per year versus OpenAI's 3x, with $447.5B in commitments through 2030.
  • Google Antigravity evicts prompt caches after a 15-second pause, making interactive coding a worst-case workload.

Google answered the “where’s the fast frontier model” question this morning with Gemini 4 Neon: a frontier-level model priced at $3 per million input tokens, $15 per million output tokens, with a 1M-token context window. It scores 61 on the Artificial Analysis Intelligence Index at its xhigh thinking level and ranks #2 on the Vals Index at 70.9%. For anyone who has watched Gemini’s flagship models get slower and pricier, Neon is the counterweight: fast, cheap, and aimed squarely at enterprise workloads.

But the more interesting reading today is the kernel.dev teardown of Gemini 4 Carbon. Carbon is unacceptably slow at real coding tasks - 22m21s on a bug fix that Opus 5.5 finished in 9m33s - and the analysis shows exactly why. It’s not the harness, and it’s not a reasoning-level mismatch. Carbon just does vastly more work per task. That distinction matters if you’re deciding which model to put in front of your CI pipeline.

A 15-second pause in Antigravity evicts your prompt cache, turning a discounted 90% cached turn into a full-price re-read of 16,000+ tokens.

Gemini 4 Neon: the fast frontier model Google actually shipped

Source: blog.google ↗

Gemini 4 Neon is Google’s answer to the complaint that frontier models are too slow and too expensive for production. At $3/M input and $15/M output, it undercuts most of the frontier tier while landing at #3 overall on the Artificial Analysis Intelligence Index (61 at xhigh thinking) and #2 on the Vals Index (70.9%).

The enterprise benchmarks are where it gets interesting. Neon ranks #1 on Legal Research Bench, Tax Agent Bench, and Zapier’s AutomationBench (56.3%), with a low fabrication rate. That’s the profile you want for agentic workflows that touch customer data - the model is cheap enough to call per-turn and accurate enough that you don’t need a human reviewing every output. Cached input drops to $0.15/M, which matters if you’re feeding it long system prompts or codebases.

It rolls out today in the Gemini API, Google AI Studio, and for paid subscribers in Google Antigravity and the Gemini app. If you’ve been running a small model for cost and a big one for quality, Neon is the candidate that collapses that split.

Genie 4: world models you can edit mid-session

Source: blog.google ↗

Google also opened Genie 4 to eligible Google AI Ultra subscribers in Project Genie. This is the world-model play: generate interactive 3D environments from text, then edit them in real time by typing prompts while exploring.

Genie 3 updates extend sessions to 3 minutes per generation (up from 60 seconds), add 360-degree panorama image grounding, and support real-time in-world events. Genie 4 pairs visuals with native audio and lets you re-prompt mid-session to change the environment on the fly. A new Project Genie agent orchestrates multimodal models to assemble these worlds from natural-language conversations.

For engineers, the interesting bit is the orchestration layer. Genie 4 isn’t one model - it’s a pipeline of multimodal models coordinated by an agent. That’s the same pattern as every other agentic product shipping this year, but applied to a domain (interactive world generation) that was demo-only six months ago.

OpenAI's models are now designing chips, and cutting area 10-13%

Source: openai.com ↗

OpenAI’s chip-design work is no longer a research curiosity. Design changes proposed by its models cut the circuit area of key parts of the Jalapeño AI chip by an estimated 10% to 13%. With AMD, the same approach delivered up to 14% improvements on flagship GPU designs.

This is the quiet story under all the model-launch noise: the models are now good enough to propose circuit-level changes that human engineers accept. That’s a compounding advantage - better chips run the models that design better chips. If you’re building on AMD or Google TPUs, the roadmap just got shorter.

Gemini 4 Carbon: slower because it does more, not because it's broken

Source: kernel.dev ↗

The kernel.dev analysis of Gemini 4 Carbon is the most useful engineering read of the day. On a real coding task, Carbon took 22m21s where Opus 5.5 took 9m33s. The slowness splits into slower inference (16-30% extra time vs Astra, 45-49% vs Opus) and dramatically longer trajectories: Carbon takes about seven times as many steps as Astra and eleven times as many actions, processing more than sixteen times as much input.

The authors ruled out harness and reasoning-level mismatches. The extra time comes from Carbon doing significantly more work per task - more steps, more actions, more tokens consumed. That’s a different failure mode than “slow inference.” It means Carbon is exploring more of the solution space, and if your tasks are well-scoped, that exploration is wasted. If your tasks are genuinely hard and underspecified, the extra steps might be exactly what you need. Know which you’re running before you pick the model.

Mid-conversation routing is a ski-rental problem

Source: edgee.ai ↗

Edgee’s blog post reframes LLM routing as a classic ski-rental problem: when should you switch models mid-session, given that switching throws away your warm cache? A router that picks the right model per turn but ignores the cache pays full price to rebuild it every time.

The numbers make the point concrete. On Anthropic, caching a prefix costs 1.25x the normal input price, and reading it back costs a tenth. So a model change mid-session can cost more than many turns of price difference between models. Edgee’s Agent Gateway treats routing as both a tier map (which model per difficulty) and a switching policy, with Self-Learning Routing replaying past sessions to recommend the cheapest capable combination. If you’re building a multi-model agent, this is the cost model you’re missing.

Antigravity eats quota because its cache dies in 15 seconds

Source: asubedi.com.np ↗

This investigation into why Google Antigravity burns through weekly quota is a masterclass in binary-searching a black box. The author disassembled the 193 MB Go binary (v2.33.0), discovered a dormant keep-alive engine, and narrowed the cache eviction cliff down to a 15-second window.

The mechanics: Gemini 4 models rely on implicit caching, where cached tokens get a 90% discount but cache misses pay 100% on the entire prefix. Even a 5-word prompt starts with 13,800 to 16,800+ tokens of system instructions, MCP schemas, and skills. If your code-review pause exceeds the TTL, every turn is a full-price cache miss. The workaround: refresh the cache just before expiry, but that only works for 3 pings. And disabling unused MCP servers and plugins shrinks the cold-turn penalty. For anyone running Antigravity or similar implicit-cache agents, this is the difference between a 10% discount and a 100% bill.

OpenAI's research acceleration: agents do half the work, humans still own the rest

Source: openai.com ↗

OpenAI’s update on research acceleration reports that its agents now complete nearly half of two-to-four-hour research tasks on their own. Agent use is surging, and the heaviest users are pulling further ahead. But humans are still central to AI development - and researchers are devoting increasing effort to safety and alignment.

The “heaviest users pull ahead” pattern is the one to watch. It suggests agentic workflows are a skill, not a toggle. Teams that invest in building the right agent harnesses and evaluation loops are compounding, while teams that treat agents as a drop-in replacement are falling behind.

The compute spending race: Anthropic 4x, OpenAI 3x

Source: epoch.ai ↗

Epoch’s analysis of compute spending shows Anthropic’s spending increasing 4x per year, outpacing OpenAI’s 3x. Anthropic’s S-1 disclosed $447.5 billion in spending commitments from September 2026 through 2030 (~$100B/year). OpenAI’s figures come from media reports and Greg Brockman’s May 2026 testimony of $50 billion expected compute spend this year.

The gap matters for anyone betting on model quality trajectories. Anthropic is spending more, faster, on the hardware that trains the models you’ll be running next year. The conversion of period-based totals into annualized run rates via log-linear regression is worth reading if you want to sanity-check the numbers yourself.

Quick hits

Source: perplexity.ai ↗

  • Perplexity’s Context API extracts full context from up to 10 URLs at once, returning clean Markdown, links, and relevant passages with freshness and source validation. Agents get page metadata (title, author, date) on every result.
  • Prime Intellect’s Model Factory is a control plane that unifies RL, SFT, pre-training, and inference, billed per GPU-hour or with dedicated Slurm/Kubernetes access. Already deployed for finance automation, computer use, scientific discovery, and chip design.
  • Google Labs’ Stitch Loop is a long-running product agent that connects to a codebase and live metrics, runs evergreen loops (Suggestions, Solutions, Build, Reflect), and opens PRs for approval. Nothing merges without your review.
  • Meta’s technique propagation agents handle 60%+ of ML technique propagation across ads ranking models since March 2026, with 3x engineering productivity, a median 4x reduction in training compute, and $45M+ in compute savings by year-end.
  • ResolveAI’s SRE-1 is a 27B model trained to find the root cause of production alerts, available only through ResolveAI’s Primary On-call product. The argument: general-purpose models fail at production because the answer exists only in the running system, not in public training data.
  • OpenAI’s Codex sandbox lets the coding agent run commands on your machine within file and network limits you control.

And two pieces worth your weekend: Zvi’s post on the METR report, arguing there’s a substantial (if under-50%) chance of superintelligence in 2027, and The Silent Consensus, an OpenAI employee’s Harvard lecture on “Black October” - critical internet infrastructure vulnerabilities found by frontier AI systems. Both argue the timeline is shorter than most engineers are planning for, and neither is a comfortable read.

Get the brief

Liked this one? Pick the beats you care about and get just those, five minutes a morning. Free, no hype, one click to leave.

About Me Author

My name is

BriefTechNews

A daily digest of what actually moved in AI, tech, crypto and fintech, assembled and written with AI, with every item linked to its source. Read More
Tags

You May Also Like