Vol. I · No. 38 Weekly Edition October 3, 2026

A fruit fly’s complete nervous system goes public. Sonnet 5.5 ties Opus 5.5 on knowledge work at half the price. Google’s Argon reaches cyber defenders without guardrails first. OpenAI ships dots and a cheaper Sol. Google tests AI compute in orbit.

Inside

The MaleCNS fruit fly connectome. Claude Sonnet 5.5 pricing and benchmarks. Gemini 4 Argon and the Fairwind Program. Project Suncatcher. OpenAI DevDay 2026.

Edited By

David Borish
From New York

Filed

Five articles, one week, sourced from The AI Spectator

Neuroscience · Connectomics No I

The AI Spectator, October 2, 2026

A whole fly
nervous system,
downloadable.

Janelia, Cambridge and Google Research mapped 166,700 neurons in one male fruit fly and released the data to researchers worldwide. It allows the first synapse-level comparison between the sexes.

By David Borish October 2, 2026

On September 3, Cell published the complete connectome of the male fruit fly central nervous system. The map covers the brain, both optic lobes and the ventral nerve cord, the insect counterpart of the spinal cord. The published abstract counts 166,700 neurons, all proofread and annotated, sorted into 11,710 types. Google Research puts the synaptic connections at about 125 million and calls it the largest brain map to date by neuron count. Janelia says the data and the tools to explore it are free to researchers worldwide.

The map is called MaleCNS. Janelia’s FlyEM team led it with the MRC Laboratory of Molecular Biology, the University of Cambridge and Google Research. Janelia imaged the tissue with focused ion beam scanning electron microscopy. Google’s software segmented the images, tracing each neuron through the three-dimensional volume. Expert teams at Janelia and Cambridge then proofread the tracing and labeled the neurons, connections and cell types.

The project began under doubt. Gerry Rubin assembled the team in 2008, when the only finished connectome belonged to the roundworm, whose 302 neurons had taken more than a decade of manual work. Rubin recalls critics who expected a fly wiring diagram to be an uninterpretable tangle. His early estimate for proofreading computer vision output was 2,000 person-years.

Computer vision closed the gap. Viren Jain worked at Janelia from 2011 to 2014, then moved to Google and joined its Connectomics team, where the fly was the first project. The flood-filling network starts at one pixel and uses a convolutional neural network to find every other pixel in the same neuron. Google says its methods cut proofreading cost and effort by a factor of 20. Janelia says imaging and AI together reduced the resources needed for a whole-brain connectome by more than 1,000-fold. Both figures come from the institutions that built the tools.

Until now every fruit fly brain connectome came from a female. The abstract reports 8,069 neuron types shared by both sexes, 138 present in both but wired differently, 289 found only in the male and 71 found only in the female. Sex-specific and dimorphic neurons concentrate in higher brain centers, while the sensory and motor periphery is largely alike.

The counts moved between versions. The earlier abstract on Google’s publication page gives 166,691 neurons and 11,691 types. The Cell version is the one to cite, under DOI 10.1016/j.cell.2026.08.015.

Because the dataset includes the nerve cord, researchers can follow a pathway from sensory neuron to motor neuron inside one dataset. Greg Jefferis of the LMB and Cambridge described it as getting from eyes to legs in one go. Three companion papers released the same day apply the map to vision, taste and social behavior.

The day after release, a Georgia Tech graduate student posted a demonstration in which the connectome drove a character in Minecraft. 404 Media, as relayed by AI Weekly, reported millions of views within a day. A daily.dev summary describes such demos as approximations that add guessed signal strengths and trained translators to the wiring. A connectome records which neurons contact which. It carries no biochemistry.

The next full brain is the larval zebrafish, in a collaboration with Harvard. Google says mapping the 86 billion neurons of a human brain is not yet possible. The largest human sample the team has reconstructed holds 50,000 cells and 150 million synapses, about 1.4 petabytes of data.

A connectome records which neurons contact which. It carries no biochemistry, and this one covers a single male fly. The AI Spectator
Read the full article →
Models · Pricing No II
70

Sonnet 5.5 ties Opus 5.5
on knowledge work at
half the price.

Anthropic released Claude Sonnet 5.5 on September 28 at $2 per million input tokens and $10 per million output tokens. Its Terminal-Bench 4.0 score is 70.6%, against 10.3% for Sonnet 5.

The headline figure on the launch page is 70.6% on Terminal-Bench 4.0, an agentic coding evaluation. Opus 5.5 scored 66.4% at its highest effort setting. Anthropic says Sonnet 5.5 generates output more than 30% faster than Sonnet 5 and costs up to 30% less per task, with per-token prices unchanged.

On GDPval-AA, which tests tasks across 44 occupations, Sonnet 5.5 scored 1844 Elo against 1846 for Opus 5.5 and 1449 for Sonnet 5. GPT-6 Sol scored 1487. On AA-Briefcase, a new long-horizon knowledge-work benchmark, the Claude figures were 1811, 1822 and 1359, with 1483 for GPT-6 Sol.

Opus 5.5 is priced at $4 per million input tokens and $20 per million output tokens, twice the Sonnet rate. That is the basis for the headline. Anthropic says Opus 5.5 remains clearly stronger at complex work that requires sustained judgment.

Effort setting matters. On FrontierCode 1.1, Sonnet 5.5 scored 52.1% at Xhigh effort and 46.2% at Max. A footnote attributes the drop to more frequent use of a code-review skill that splits work across subagents, which led to a timeout or out-of-scope edits in two cases Cognition examined. At High effort, Anthropic says Sonnet 5.5 matches GPT-6 Sol’s best score for about a fifth of the cost per task.

Enterprise testers reported cost per answer. Balyasny Asset Management ran 2,441 finance tasks and measured about 121,000 tokens per answer for Sonnet 5.5 against 497,000 for Sonnet 5. Base44 reported 3.6 iterations per build where Opus 5 needed 7.7.

The caveats are on the launch page. Artificial Analysis ran GDPval-AA and AA-Briefcase on a pre-release deployment with a structured-output bug, which Anthropic says likely understates Sonnet 5.5. The Claude models are compared at different effort settings on Terminal-Bench, and the page does not explain what held Sonnet 5 back at 10.3%. Tester task suites are private, and the testers have commercial ties to Anthropic.

Developers running Sonnet with thinking off need to switch to a setting called between_tools before upgrading. Sonnet 5.5 is also the first Sonnet with classifiers meant to block reasoning extraction. The practical step is to run your own workloads at Low and Medium effort before adopting higher settings.

70.6%
Terminal-Bench 4.0
Sonnet 5: 10.3%
1,844
GDPval-AA Elo
Opus 5.5: 1,846
$2 / $10
Per Million Tokens
In / Out, Half of Opus 5.5
Read the full article →
// security :: cyber_defense :: access_control No III
[ SEQUENCE LOG, SEPTEMBER 30, 2026 ]

Google’s new model
goes to cyber defenders
without guardrails first.

$ announce --model "gemini-4-argon" --date 2026-09-30
> First access: trusted cyber defenders, Fairwind Program
$ price --unit per_million_tokens
> input: $2 • output: $10 • cached input: 95% off input rate
$ limits --output_tokens
> 1,000,000 (previous: 64,000)
$ release --general
> Paid API and Google AI Ultra: date not given

Google announced Gemini 4 Argon on September 30. Koray Kavukcuoglu’s post describes a model built to sustain deep reasoning across long, multi-step work in software engineering, enterprise knowledge work and cyber defense. The first access group is trusted defenders in the Fairwind Program. Paid API customers and Google AI Ultra subscribers follow, with no date beyond as soon as possible. Google says it is engaged in the U.S. government’s voluntary process for pre-release model access.

Launch pricing is $2 per million input tokens and $10 per million output tokens, with cached input at 95 percent off, about ten cents per million. The output limit rises to 1 million tokens from 64,000. The post gives no data on how the larger window affects accuracy or cost.

Google trained Argon to find, validate and patch critical vulnerabilities, and will release it without cyber guardrails to trusted defenders and its internal teams. Wiz is using it through Scan for Good, and Google says Argon found a critical flaw in healthcare software used by hospitals worldwide that earlier frontier models missed. Google does not name the software or say whether the flaw is fixed. On CWE-bench v1, Argon ties for first at 68 percent.

Gated release is a dependency. The no-guardrail version goes to a group Google selects, and everyone else waits. The Open-Prem Inflection Point V3 asks organizations to price that kind of dependency when they choose between hosted and self-hosted models.

Internal results. In libgav1, Google’s open source video decoder, Argon agents replaced 32,000 lines of SIMD code in an existing Rust port. The result runs 2.7 times faster than that port with identical video output. Rewrites of C and C++ code to Rust range from tens of thousands of lines in re2 to more than 800,000 for the Zircon kernel of Fuchsia OS, and remain under audit. Memory optimizations by Argon agents are expected to free more than 300 TiB once rolled out.

Every benchmark in the post is self-published: DeepSWE v1.1 at 77.9 percent, AutomationBench at 51.3 percent, LVBench at 91.7 percent. The internal cyber benchmarks give no scores and cannot be checked from outside. Read the charts against the comparison models before citing relative performance.

// DATA_LOG
68%
CWE-BENCH V1
Tie for first on vulnerability remediation, per Google
2.7x
LIBGAV1 SPEED
Versus the Rust port, with identical video output
1M
OUTPUT TOKENS
Output limit, up from 64,000
300+ TiB
MEMORY
Expected to be freed once the optimizations roll out
Read the full article →
Space · Infrastructure No IV

Google Tests Whether
AI Compute Can Fly
in a Tight Formation

§ § §

Project Suncatcher offers proton-beam data, orbital modeling and a bench-scale laser link. Thermal management, ground links and multi-year reliability remain untested.

Solar panels in low Earth orbit generate up to eight times what the same panels produce on the ground, and they do it around the clock. That is the premise of Project Suncatcher, described in a new paper and blog post from Travis Beals of Google Research: put Tensor Processing Units on a tight constellation of satellites, link them with lasers, and run them on solar power.

The timing reflects a ground constraint. Data centers are running into power-grid limits in market after market, and new grid capacity or a gas or nuclear plant takes years of permitting. A satellite does not wait in that line. Google is aiming at that bottleneck, not at a claim that computing in space is simply cheaper.

The paper lists Beals and eight co-authors, including Google VP Blaise Agüera y Arcas, and organizes its evidence around four questions. Bandwidth. The target for satellite links is tens of terabits per second. A bench demonstrator reached 800 Gbps in each direction, 1.6 Tbps combined. Signal strength falls with the square of distance, so that speed requires satellites within a kilometer of each other or closer.

Formation. Google modeled an 81-satellite formation at 650 kilometers altitude with a one-kilometer cluster radius. The models suggest Earth’s gravity does most of the work of holding it together, with modest station-keeping. Next-nearest neighbors oscillated between roughly 100 and 200 meters apart over a single orbit.

Radiation. Google fired a 67 MeV proton beam at a Trillium v6e TPU. The most sensitive component, its onboard memory, showed trouble at 2 krad(Si), nearly three times the roughly 750 rad(Si) a shielded five-year mission is expected to accumulate. Nothing failed outright up to the maximum tested dose of 15 krad(Si).

Cost. Google extends the fall in launch prices to a projection of under $200 per kilogram by the mid-2030s. At that price it estimates a space-based data center’s cost per kilowatt-year could land close to a terrestrial one. The comparison rests on Google’s own extrapolation, not on a price anyone has paid to operate a data center in orbit.

Competitors have more flight time. Starcloud put an Nvidia H100 in orbit in November 2025, raised $170 million at a $1.1 billion valuation, and filed with the FCC for a constellation of up to 88,000 satellites. SpaceX, which acquired xAI in February 2026, has filed for a distributed-computing network of its own. Jeff Bezos expects gigawatt-scale orbital data centers within roughly a decade. Eric Schmidt acquired Relativity Space in 2025 and is steering it toward the same goal.

Google’s two Planet-built prototypes fly in early 2027. Beals says thermal management in a vacuum has not had a bench demonstration, and neither has ground-to-orbit communication at workload bandwidth or reliability over a multi-year mission. Those are the harder parts of the problem, and they trail the results reported so far.

The AI Spectator • October 3, 2026 • Space Infrastructure
Read the full article →
Economics · Agents · Enterprise No V

Sol matches
Astra at one-fifth
the price.

OpenAI reports that GPT-6.1 Sol matches GPT-6 Astra on DeepSWE v1.1 at $2 and $10 per million tokens, against $10 and $50. Dots, agents with their own cloud computer, lead a list of more than 20 DevDay announcements.

The figure appeared on September 29 in OpenAI’s DevDay 2026 recap. The company also says the ChatGPT platform reaches a collective 1.2 billion weekly users. Availability differs sharply from item to item, so plan tier and release status matter as much as the headline.

The lead launch is dots, agents that run continuously on GPT-6 Astra with their own cloud computer and browser and access to more than 4,000 apps through plugins. They roll out on Pro and Business Premium plans in eligible markets. Enterprise, Edu and Healthcare workspaces get a beta that stays off until an administrator enables it. An idle dot does read-only research. Custom Rules let a user allow, require approval for or block actions. OpenAI advises reviewing consequential work, and its examples come without a measured success rate.

Specialist dots are limited to enterprise pilots and receive their own identity, credentials and IT-provisioned hardware. OpenAI names procurement, invoice processing, email marketing, customer support and commercial contracting as areas tested internally.

The cost claims carry caveats. On OSWorld 2.0’s offline set, Sol lands within 2.1 points of Astra at about one-seventh of the cost per task. On Terminal-Bench Science 0.1, Sol costs $5.47 per task at maximum effort against $23.21 for Opus 5.5 and $23.80 for Astra. Astra still posts the highest score at 68.1 percent, and OpenAI recommends it for the hardest research tasks. A footnote says the cost shown for Claude Fable 5.1 on AutomationBench omits fallbacks, which occurred on about 40 percent of tasks.

Outside numbers are thin. Artificial Analysis, as summarized by Latent Space’s AINews, places Sol one point below Astra on its Intelligence Index at $0.72 per task against $3.26, and reports that Sol uses 10 to 30 percent more output tokens than GPT-6 Sol.

Speed has its own tier. Ultrafast offers up to 8 times faster generation in Codex, about 300 tokens per second. Astra Ultrafast is on the new Pro 500 plan, which carries 25 times the Plus allowance, and on Enterprise.

Two commercial moves complete the list. The OpenAI Marketplace lets eligible enterprise customers apply part of an existing OpenAI commitment toward partner software from a first group of 32 vendors. Sign in with ChatGPT lets Plus and Pro users spend plan allowance across 16 partner tools. Bedrock Managed Agents, built with Amazon, runs OpenAI agents entirely inside AWS.

Read the full article →
Key Figures
$2 / $10
GPT-6.1 Sol price per million input and output tokens, against $10 and $50 for GPT-6 Astra
1.2B
Collective weekly users across the ChatGPT platform, per OpenAI
4,000+
Apps a dot can reach through plugins
$5.47
Sol cost per task on Terminal-Bench Science 0.1 at maximum effort, against $23.80 for Astra, which still scores highest at 68.1 percent

The AI Spectator Weekly is published at davidborish.com/the-ai-spectator

Frameworks explored this issue:
Open-Prem Inflection Point V3  ·  The Exponential Replacement Curve

Vol. I · No. 38 · October 3, 2026 · Edited by David Borish · New York