Anthropic
Claude and Claude Code; frontier models and agent workflows.
San FranciscoOfficial site ↗
AFFLATUS / AI OBSERVATORY
The world building AI
The companies, models and turning points shaping the next era of intelligence. A global field guide, grounded in evidence.
CAPABILITY · CAPITAL · CONTROL
Explore the story ↓A stronger model is only the beginning. Compute, distribution and trust determine what follows.
Dots trace land. Colored markers locate selected headquarters.
01 / 08THE ECOSYSTEM
Intelligence is a global supply chain
The industry is bigger than the race between model labs. Advanced chips depend on fabrication and memory; clusters need networking, electricity and cooling. Agents need data, tools and distribution. Creative systems, robotics and drug discovery extend the frontier into different worlds.
An editorial map of critical industry nodes, not an exhaustive register or a claim that every company is irreplaceable. Primary categories simplify companies that span several layers.
100selected organizations
14industry layers
13HQ markets
Follow the dependencies
Drag to rotate · markers are headquarters, not compute capacity
OpenAI
GPT, Codex and the API platform.
San FranciscoOfficial site ↗
Microsoft
Azure cloud; investor in OpenAI and Anthropic
RedmondOfficial site ↗
Alphabet (Google)
Gemini, Google Cloud and TPUs
Mountain ViewOfficial site ↗
Amazon
AWS and Trainium
SeattleOfficial site ↗
Meta
Muse Spark models
Menlo ParkOfficial site ↗
Read the full directory (100)
Frontier models
The intelligence layer: closed services and open weights.
Anthropic
Claude and Claude Code; frontier models and agent workflows.
San FranciscoOfficial site ↗
OpenAI
GPT, Codex and the API platform.
San FranciscoOfficial site ↗
Alphabet (Google)
Gemini, Google Cloud and TPUs
Mountain ViewOfficial site ↗
Meta
Muse Spark models
Menlo ParkOfficial site ↗
SpaceX (incl. xAI)
Grok models, Colossus compute, Cursor
StarbaseOfficial site ↗
Alibaba
Qwen models and Alibaba Cloud
HangzhouOfficial site ↗
Baidu
Ernie models
BeijingOfficial site ↗
DeepSeek
Open-weights models and efficient inference.
HangzhouOfficial site ↗
Moonshot AI
Kimi models and long-context agents.
BeijingOfficial site ↗
Zhipu (Z.ai)
GLM models and developer APIs.
BeijingOfficial site ↗
MiniMax
Language, voice and video generation.
ShanghaiOfficial site ↗
Mistral AI
European models and enterprise AI deployment.
ParisOfficial site ↗
Cohere
Enterprise models, retrieval and private deployment.
TorontoOfficial site ↗
Xiaomi
MiMo models and device AI.
BeijingOfficial site ↗
Tencent
Hunyuan models and consumer distribution.
ShenzhenOfficial site ↗
ByteDance
Seed models, Doubao and multimodal products.
BeijingOfficial site ↗
StepFun
Step models and multimodal reasoning.
ShanghaiOfficial site ↗
Sarvam AI
Language models for India and local-language access.
BengaluruOfficial site ↗
Upstage
Solar models and document intelligence.
SeoulOfficial site ↗
Cloud & distribution
Access, training capacity and routes to customers.
Microsoft
Azure cloud; investor in OpenAI and Anthropic
RedmondOfficial site ↗
Amazon
AWS and Trainium
SeattleOfficial site ↗
Oracle
Oracle Cloud infrastructure and database distribution.
AustinOfficial site ↗
G42
Regional AI infrastructure and deployment.
Abu DhabiOfficial site ↗
Accelerators
GPUs, custom silicon and inference systems.
NVIDIA
GPUs and systems
Santa ClaraOfficial site ↗
Broadcom
Custom accelerators and data-centre networking.
Palo AltoOfficial site ↗
AMD
Instinct GPUs
Santa ClaraOfficial site ↗
Marvell
Custom silicon and optics DSPs
Santa ClaraOfficial site ↗
Arm
CPU architecture in AI servers
CambridgeOfficial site ↗
Cerebras
Wafer-scale compute and inference systems.
SunnyvaleOfficial site ↗
Intel
Server CPUs and foundry
Santa ClaraOfficial site ↗
Qualcomm
On-device AI chips
San DiegoOfficial site ↗
Huawei
Ascend accelerators and the domestic compute ecosystem.
ShenzhenOfficial site ↗
Cambricon
AI processors and software for domestic deployment.
BeijingOfficial site ↗
Preferred Networks
AI accelerators and industrial deep-learning systems.
TokyoOfficial site ↗
Chip design & manufacturing
EDA, fabrication, packaging, testing and the equipment behind them.
TSMC
Leading-edge foundry for AI chips
HsinchuOfficial site ↗
ASML
EUV lithography
VeldhovenOfficial site ↗
Applied Materials
Deposition and etch tools
Santa ClaraOfficial site ↗
Lam Research
Etch tools for memory and logic
FremontOfficial site ↗
KLA
Process control and inspection
MilpitasOfficial site ↗
SMIC
Domestic logic-chip fabrication.
ShanghaiOfficial site ↗
NAURA
Semiconductor manufacturing equipment.
BeijingOfficial site ↗
AMEC
Etch and deposition equipment for chip manufacturing.
ShanghaiOfficial site ↗
Tokyo Electron
Semiconductor production equipment.
TokyoOfficial site ↗
Advantest
Test systems for advanced semiconductors.
TokyoOfficial site ↗
Synopsys
EDA, semiconductor IP and chip-to-system verification.
SunnyvaleOfficial site ↗
Cadence
EDA, verification, multiphysics and chiplet design tools.
San JoseOfficial site ↗
ASE Technology
Advanced semiconductor packaging and testing.
KaohsiungOfficial site ↗
Amkor Technology
Advanced packaging, chiplet integration and semiconductor testing.
TempeOfficial site ↗
Memory & storage
HBM, DRAM, flash and durable storage.
SK hynix
HBM and advanced DRAM for accelerators.
IcheonOfficial site ↗
Micron
HBM, DRAM and SSD
BoiseOfficial site ↗
Samsung Electronics
HBM, DRAM, NAND and foundry services.
SuwonOfficial site ↗
Sandisk
NAND flash
MilpitasOfficial site ↗
Western Digital
Data-centre hard drives
San JoseOfficial site ↗
Seagate
Data-centre hard drives
DublinOfficial site ↗
Networking
Moving data across the cluster.
Arista Networks
AI data-centre switches
Santa ClaraOfficial site ↗
Cisco
Networking
San JoseOfficial site ↗
Coherent
Optical transceivers
SaxonburgOfficial site ↗
Lumentum
Lasers and optical components
San JoseOfficial site ↗
Credo
Active electrical cables
San JoseOfficial site ↗
Astera Labs
PCIe/CXL connectivity
San JoseOfficial site ↗
Ciena
Data-centre interconnect
HanoverOfficial site ↗
Servers & GPU clouds
The infrastructure that turns chips into available compute.
Dell Technologies
AI servers
Round RockOfficial site ↗
Supermicro
AI servers
San JoseOfficial site ↗
CoreWeave
GPU cloud
LivingstonOfficial site ↗
Nebius
GPU cloud
AmsterdamOfficial site ↗
GDS
High-performance data-centre infrastructure.
ShanghaiOfficial site ↗
Energy & cooling
Electricity, grid connections and thermal limits.
Vertiv
Power and liquid cooling
WestervilleOfficial site ↗
GE Vernova
Gas turbines and grid equipment
Cambridge, MAOfficial site ↗
Constellation Energy
Nuclear power for data centres
BaltimoreOfficial site ↗
Vistra
Power generation
IrvingOfficial site ↗
Eaton
Electrical equipment
DublinOfficial site ↗
Oklo
Small modular reactors
Santa ClaraOfficial site ↗
Coding agents
From writing a line to owning a development task.
Cursor
Agent development inside the editor and in the cloud.
San FranciscoOfficial site ↗
Cognition
Devin: delegated software engineering tasks.
San FranciscoOfficial site ↗
Replit
Build and deploy applications with an agent.
Foster CityOfficial site ↗
GitHub
Copilot and the repository-to-pull-request workflow.
San FranciscoOfficial site ↗
Data & developer tools
Datasets, evaluation, orchestration and retrieval.
Hugging Face
Shared models, datasets and open ML tooling.
New YorkOfficial site ↗
Databricks
Lakehouse data infrastructure and enterprise AI.
San FranciscoOfficial site ↗
Scale AI
Training data, evaluation and enterprise deployment.
San FranciscoOfficial site ↗
LangChain
Agent orchestration, tracing and evaluation.
San FranciscoOfficial site ↗
Pinecone
Vector retrieval for grounded AI applications.
New YorkOfficial site ↗
Enterprise & search
Workflow integration and useful answers.
Palantir
AI platforms for enterprise and government
DenverOfficial site ↗
Salesforce
Enterprise AI agents
San FranciscoOfficial site ↗
ServiceNow
Workflow AI
Santa ClaraOfficial site ↗
Snowflake
Data platform
BozemanOfficial site ↗
Perplexity
Search, answer synthesis and research workflows.
San FranciscoOfficial site ↗
Creative & voice AI
Image, video, sound and interactive speech.
Adobe
Creative AI
San JoseOfficial site ↗
Runway
Video generation and world models.
New YorkOfficial site ↗
ElevenLabs
Speech synthesis and conversational voice agents.
LondonOfficial site ↗
Black Forest Labs
FLUX image generation and editing.
FreiburgOfficial site ↗
Midjourney
Creative image generation.
San FranciscoOfficial site ↗
Stability AI
Open generative media models.
LondonOfficial site ↗
Physical AI
Intelligence that must work in the physical world.
Figure
Humanoid robots and embodied AI.
San JoseOfficial site ↗
Unitree
Quadruped and humanoid robot hardware.
HangzhouOfficial site ↗
Tesla
Autonomous driving and humanoid robotics.
AustinOfficial site ↗
Physical Intelligence
General-purpose robot foundation models.
San FranciscoOfficial site ↗
Science & evaluation
Drug discovery and tests of genuine generalization.
Isomorphic Labs
AI-driven drug discovery.
LondonOfficial site ↗
Recursion
Biological data and computational drug discovery.
Salt Lake CityOfficial site ↗
METR
Independent evaluation of autonomous task capability.
BerkeleyOfficial site ↗
ARC Prize
Benchmarks for generalization to unfamiliar tasks.
San FranciscoOfficial site ↗
02 / 08THE MODEL FRONTIER
Which models deserve attention?
A shortlist of leading families and meaningful efficiency alternatives. Superseded generations and weak entries are omitted. Each row preserves the measured configuration: an effort setting or fallback can change both capability and cost.
Model observations: 9 October 2026 · Artificial Analysis Intelligence Index v4.3.2 · displayed, rounded scores · tied scores share a rank. This is a general capability index, not a coding-agent leaderboard.
NEW · 7 OCTOBER 2026
Claude Haiku 5.5
Intelligence finds a smaller form.
A fast helper deserves its own settings. Compare the wait, cost and capability before choosing how much thinking a task needs.
Source ↗Read our analysis ↓highAA index38
USD / AA task$0.08
First chunk · seconds28.01
Output tokens / s160
Changing effort here compares measured configurations; the main family ranking retains max. Output speed measures generation after the first chunk.
View this configuration ↗19 selected configurations
AA Intelligence Index · higher is better
Cost is USD per Artificial Analysis benchmark task, not price per million tokens. Output speed excludes the wait for reasoning; first-chunk latency makes that delay visible. Missing values stay missing. Open weights does not imply an unrestricted licence.
Artificial Analysis · model leaderboardCapability and the cost of thinking
Hover, focus or select a point to inspect the model.
Colors identify makers; names and official marks retain identity. Marker size is equal. Dashed line: capability–cost frontier within this filtered selection.
03 / 08US & CHINA
Two powers. One interdependent world.
America combines frontier labs, hyperscale clouds and deep capital markets. China combines a large application market, a broad model ecosystem and a drive toward domestic compute. Neither story can be understood without the fabrication, equipment and memory suppliers beyond their borders.
The balance changes with the question
Frontier model capability
AA index points · observed 9 October 2026
Best selected model configuration: Opus 5.5 vs MiMo-V2.6-Pro. A 12-point index gap is not a percentage or a measure of national scientific capability.
Artificial Analysis · model leaderboardUnited States
Strengths: leading selected model scores, GPU design, cloud distribution and private funding. Watch: infrastructure economics, concentration and the reliability of deployed agents.
China
Strengths: competitive open weights, efficient inference, local distribution and industrial deployment. Watch: access to advanced hardware and whether model adoption converts into durable business value.
These are editorial interpretations. Stanford’s March 2026 model gap and its 2025 investment totals describe different periods and methods from the October leaderboard. They must not be combined into a single national “AI power” score.
Stanford · AI Index 2026Beyond a single leaderboard
A research snapshot and editorial reading of the competing strengths.
Research and industrial deployment
United States
AI Index 2026 reports more top-tier models and higher-impact patents.
China
The report finds a lead in publication volume, citations, patent output and industrial robot installations.
These measure different parts of the ecosystem, not a single winner or the reliability of a deployed agent.
Source ↗Control over the model
United States
The leading selected OpenAI and Anthropic configurations are closed services.
China
MiMo, GLM, Kimi and Qwen Flash offer open-weight alternatives in this shortlist.
Compare the actual licence, deployment cost and operational responsibility. Neither every Chinese model nor every American model follows one approach.
Source ↗The shared hardware constraint
United States
Chip design and hyperscale infrastructure are strengths; leading AI chip fabrication remains highly concentrated at TSMC.
China
Domestic accelerators and fabrication are strategic nodes to watch, alongside access to the global equipment and memory stack.
Taiwanese fabrication, Dutch lithography and Korean memory make this a global supply-chain contest. This is a structural interpretation, not a claim of technological self-sufficiency.
Source ↗Where each ecosystem is represented
Counts within this editorial directory, ranked by industry layer. They are not market shares, capacity or national rankings. Hover or focus a layer to trace it across both sides.
- 01Accelerators7
- 02Networking7
- 03Chip design & manufacturing6
- 04Frontier models5
- 05Energy & cooling5
- 06Data & developer tools5
- 07Enterprise & search5
- 08Coding agents4
- 01Frontier models10
- 02Chip design & manufacturing3
- 03Accelerators2
- 04Servers & GPU clouds1
- 05Physical AI1
- 06Cloud & distribution0
- 07Memory & storage0
- 08Networking0
04 / 08ANTHROPIC · BEFORE THE IPO
The frontier meets the balance sheet
Claude’s capability is one part of the story. Before a public listing, the more revealing questions concern revenue quality, infrastructure obligations, customer concentration and the rights that come with ownership.
IPO watch · reported confidential filing
Reuters reports a confidential filing and a possible post-midterm listing. A public prospectus, final price, ticker and listing date were not verified in this snapshot. Reported ambitions above $2 trillion are not a priced public valuation.
Reuters · reported IPO prospectusSeries H post-money valuation
$65B raised. A private financing valuation, not a market capitalization.
Source ↗Revenue in 2025
Approximately 12× year-on-year. Keep separate from the $47B annualized run-rate announced in May 2026.
Source ↗Source ↗Infrastructure commitments
About 80% is non-cancelable. These are multiyear contractual commitments, not a single year’s spending or current debt.
Source ↗What makes the story compelling
Frontier capability and a developer workflow meet enterprise distribution through major clouds. The opportunity is to make Claude a dependable layer inside software, rather than rely only on a consumer subscription. That is an editorial thesis, not a revenue forecast.
Anthropic · Series HWhat can break the thesis
Large fixed compute commitments require sustained demand. Reuters reports two unnamed customers represented roughly one quarter of 2025 revenue. Model leadership alone does not establish pricing power, retention or attractive margins.
Reuters · reported IPO prospectusReuters · compute commitmentsPartners are also investors and suppliers
AWS is the primary cloud and training partner; Claude is also distributed through Google Cloud and Microsoft Foundry. Announced gigawatt-scale capacity plans still require delivery and utilization.
Anthropic · Series HA different governance structure
Anthropic describes itself as a Public Benefit Corporation. Stockholders and the Long-Term Benefit Trust elect its board. The public filing must clarify the protections and economic rights that new shareholders will actually receive.
Anthropic · company & governanceWhat to read when the public filing arrives
Revenue quality & cloud relationships
Look for recognized revenue, retained customers, gross margin, revenue sharing and related-party transactions. Cloud investors may also be suppliers and distribution partners.
Reuters · reported IPO prospectusLosses, cash and accounting charges
Reuters reports an approximately $42B 2025 net loss, including about $34B in financing fair-value charges, and an operating loss above $8B. Net loss is not equivalent to cash burn. Read the cash-flow statement and reconciliations.
Reuters · reported IPO prospectusCompute obligations & delivery risk
Check payment schedules, termination rights, utilization, supplier concentration and the gap between announced capacity and operating capacity.
Reuters · compute commitmentsGovernance, dilution & safety
Read the actual voting rights, share classes, public-benefit obligations, board protections, dilution and risk factors in the public prospectus. Do not assume that common ownership means control.
Reuters · reported IPO prospectus05 / 08AGENTIC CODING
From a suggestion to a collaborator
We treat 2025 as the first mainstream year of agentic coding, with 2024 as its prelude. The milestones below track changes in the unit of work: snippets, repository tasks, cloud delegation, parallel agents and longer autonomous runs. This is a curated history of consequential changes.
Devin puts the agent in a workspace
A task-oriented system combines planning, terminal, editor and browser. The precursor to delegating development rather than requesting snippets.
Devin puts the agent in a workspaceMCP opens a tool-connection layer
A common protocol makes external data and tools accessible to models. It helps agent integrations move beyond one-off connectors.
MCP opens a tool-connection layerDevin becomes generally available
Delegated software engineering becomes a purchasable workflow; availability is distinct from guaranteed autonomous success.
Devin becomes generally availableCopilot adds agent mode
The IDE assistant can iterate on edits and commands, bringing tool use into the developer’s everyday environment.
Copilot adds agent modeClaude Code enters research preview
The terminal becomes an agent interface, with repository context, file edits and command execution.
Claude Code enters research previewCodex CLI brings the agent to the terminal
An open-source local interface gives reasoning models access to repository code and terminal tools. Local work joins the emerging cloud workflow.
OpenAI · Codex CLI launchCursor moves tasks into the background
Background agents enter preview in separate remote environments. Developers can monitor, redirect or take over asynchronous repository tasks.
Cursor · background agents previewCodex delegates work to cloud sandboxes
Parallel isolated tasks turn the coding agent into an asynchronous collaborator. Results still need review and testing.
Codex delegates work to cloud sandboxesCopilot takes issues through to PRs
Repository tasks connect to the pull-request review loop. The unit of work becomes an issue, rather than a completion.
Copilot takes issues through to PRsClaude 4 and Claude Code GA
Longer tool workflows and a generally available command-line product make agent coding more broadly accessible.
Claude 4 and Claude Code GAGemini CLI brings an open-source agent
An open-source command-line agent broadens the ecosystem beyond proprietary editors and cloud services.
Gemini CLI brings an open-source agentGPT-5-Codex specializes for agent coding
An agent-oriented model emphasizes extended coding and review workflows, alongside improvements to the coding product.
GPT-5-Codex specializes for agent codingSonnet 4.5 pushes sustained coding
The focus shifts toward longer software tasks and agent scaffolding. Vendor demonstrations remain distinct from independent reliability tests.
Sonnet 4.5 pushes sustained codingComposer and a workspace for parallel agents
Cursor 2.0 pairs a coding model with parallel agents, isolated worktrees and browser verification. Review and testing become central parts of the interface.
Cursor · 2.0 and ComposerAntigravity connects editor, terminal and browser
Google launches a public preview with an agent manager and reviewable artifacts. Work is organized around delegated tasks spanning multiple development surfaces.
Google · Antigravity launchCursor studies long-running agents
Research on planners and workers explores extended autonomous coding. A successful experiment does not establish general unattended reliability.
Cursor studies long-running agentsThe Codex app coordinates parallel work
A dedicated app organizes multiple agents, isolated worktrees and automations around a developer’s projects.
The Codex app coordinates parallel workARC-AGI-3 tests unfamiliar interactive worlds
Evaluation asks agents to explore and learn unseen environments. Coding success alone is insufficient evidence of general intelligence.
ARC-AGI-3 tests unfamiliar interactive worldsMirrorCode tests some weeks-long coding tasks
METR and Epoch AI report preliminary evidence on selected long coding tasks. Success on some tasks is consequential, but does not establish reliable autonomy across arbitrary weeks-long projects.
METR · MirrorCode preliminary resultsCursor joins SpaceX
A major agent-coding product moves into a larger model and compute ecosystem. Product names and parent companies must be distinguished.
Cursor joins SpaceXOpus 5.5 advances the coding frontier
New benchmark results broaden the frontier comparison. Different effort settings and fallback policies must remain visible.
Opus 5.5 advances the coding frontierSonnet 5.5 raises the agent benchmark
Anthropic reports 70.6% on Terminal-Bench 4.0. This is a vendor result on a defined task suite, not an AGI milestone percentage.
Sonnet 5.5 raises the agent benchmarkDevDay expands shared agent workflows
GPT-6.1 Sol and Codex workflow updates put more emphasis on shared environments, review and coordinated agent work.
DevDay expands shared agent workflowsHaiku 5.5 makes the small-task layer cheaper
A small model with adjustable effort changes how agents divide work. The useful design question becomes which tasks can be delegated, verified and escalated, rather than which model runs every step. This is our interpretation of the release.
Anthropic · Haiku 5.5 release & pricing
06 / 08TOWARD AGI
Progress is real. The finish line is undefined.
There is no scientifically agreed percentage for “distance to AGI.” This progress bar is an editorial capability map: click each gate to inspect what has been demonstrated and what remains unproven. It is neither a countdown nor an estimate of completion.
Current frontier
Supervised agents
Coding agents can iterate in real environments. Sonnet 5.5’s 70.6% Terminal-Bench 4.0 result shows progress on a specific test, with substantial failures remaining.
Terminal-Bench 4.0 · Sonnet 5.570.6%
Vendor-reported success rate
Why benchmark progress cannot be averaged into AGI
Terminal coding tasks, business automation, unfamiliar interactive games and long-horizon reliability test different abilities. Their denominators, scaffolds, access policies and success criteria differ. Reaching a high score on one does not settle transfer, continual learning or safe control.
METR · task time horizonsARC-AGI-3 · launch baselineARC Prize · current leaderboard07 / 08CONSUMER AI
Who uses AI. Who pays for it.
Traffic, spending and distribution reveal different markets.
a16z · Consumer AI, edition 7Olivia Moore · 5 October 2026
4.5%paid personal subscriptions¹
$25median payer / month
$903top 1% / month
Spending concentration
US consumer panel · August 2026 · bars scaled to group maximum
100 dots = all observed spending; 20 shaded dots ≈ the top 1% share.
How products monetize
44 AI-native web products; overlapping models. ¹ChatGPT, Gemini or Claude. Panel spending is not company revenue.
08 / 08SUBSCRIPTION ECONOMICS
The price is only the beginning.
Compare a plan, model and workload together.
SemiAnalysis · Subscription value testsAndrew Megalaa, Max Kan & Dylan Patel · 5 October 2026
Compare the reported tests
Opus 5.5 / GPT-6.1 Sol
OpenAI1×
Anthropic≈5×
Reported API-equivalent ratio; not a capability score.
API-equivalent value ≠ quality or guaranteed savings. Dated agentic-workload estimates; limits can change.
Measurement method
Isolate input, cache write, cache read and output; observe meter increments; apply weekly caps and API prices.
Source ↗AFFLATUS ORIGINAL ANALYSIS
The scarce resource is a finished task.
A subscription purchases access to a changing pool of compute. The useful question is how much of that access becomes work you can accept. We separate the reported allowance, the fraction you can use, and the quality you verify.
Evidence boundary: we read the publicly accessible SemiAnalysis article. Its paid Tokenomics Dashboard and the locked third-party-plan section were not available. The 5 October results remain attributed; we did not retest subscription limits.
01 / The denominator changes the story
An API dollar is a reference price.
API-equivalent value multiplies usable tokens by the API list price. If that price halves while token allowance stays fixed, the reported dollar value halves too. Access has not changed. Read token volume and task acceptance alongside the headline ratio; an expensive API can inflate the apparent subsidy.
Source ↗02 / Rhythm is part of the product
An allowance must be reachable.
A large weekly allowance can still be a poor fit for a compressed deadline. Model-specific caps, reset windows and bursty work affect utilization. A fair comparison should record the fraction actually consumed during your working hours, rather than assume every buyer exhausts the plan.
At a 5× reference ratio, the utilization break-even is Claude utilization = one fifth of OpenAI utilization. That is algebra under the reported premise, not a purchasing verdict.
Make the assumption visible
Start from SemiAnalysis’s approximate 5× daily-driver ratio. Move the utilization assumptions to see how much of the nominal advantage you can reach. Quality and task mix are held outside this calculation.
3.75×Utilization-adjusted API-equivalent ratio
5 × 60% ÷ 80% = 3.75×. A scenario, not measured savings.
03 / Haiku changes the division of labor
Buy precision where it matters.
Our architecture thesis: let a small model extract, classify and summarize; let a stronger model plan, resolve ambiguity and review. Escalation needs an explicit trigger such as missing evidence or a failed check. Savings survive only if verification and retries cost less than the work delegated.
Measure cost per accepted result = total model, tool, retry and review cost ÷ accepted results. The AA index helps select candidates; your own acceptance tests decide deployment.
04 / Efficiency has several frontiers
Choose the bottleneck first.
In this shortlist, MiMo scores 46 at $0.13 per AA task; Haiku max scores 43 at $0.21. Haiku’s case rests on the wider system: effort control, output speed, tool fit and operational continuity. A cost–capability plot cannot decide interactive responsiveness or verified work quality.
Source ↗A small model can generate quickly yet think for a long time at max. Compare first-chunk delay separately from tokens per second.
The shape of a request changes its bill
A fixed-token API scenario using the official price table. Adjust total prompt length and cache hits; output stays at 1,000 tokens. This is independent of subscription caps and of the AA task mix.
Haiku 5.5 price tier: up to 100k prompt tokens. USD per modeled request.
Cost = fresh input × input rate + cached input × read rate + output × output rate, divided by 1M. Above 100k, the higher Haiku rate applies to the whole request. No cache writes, tools, retries, batch discounts or tokenizer differences are modeled. Equal token counts do not imply equal task quality.
Source ↗05 / A bridge from subscribers to builders
The next moat may be a workflow.
Anthropic announced monthly API credits for Max and Team alongside this release. Our inference is that a subscription can become a route into API applications: the user who learns to compose and verify a workflow may later pay for sustained usage. Credits invite experimentation; durable margins still depend on retention, utilization and the full cost to serve.
This is a testable business thesis, not a revenue forecast. Watch accepted-task cost, repeat API use after credits, escalation rate and paid retention. Keep the October 5 subscription tests separate from the October 7 product changes.
Source ↗Read across the evidence
Our reading: adoption, willingness to pay and cost to serve are separate questions. Use the evidence wall to compare the two lenses.
Evidence, not a moving headline
Research cut-off: 9 October 2026. Data publication dates remain visible. Official releases document product claims; independent evaluations measure specific systems; reported IPO figures remain attributed to Reuters. Company classifications and interpretations are editorial.
Directory geography is approximate headquarters location. Product brands, subsidiaries and research organizations are included separately where useful; counts do not represent independent parent companies. Earlier official company and logo records were checked on 27 September; this research refresh adds the new organizations.
Citation & methodology
AFFLATUS. The world building AI / 谁在构建 AI 的世界. 2026-10-09. https://feida.au/sectors.html Snapshot: /data/sectors-observatory/2026-10-09.json Model data: Artificial Analysis Intelligence Index v4.3.2.
The model shortlist uses one representative high-capability configuration per family and adds efficient alternatives. Scores are transcribed from the displayed public leaderboard. Rankings and CSV exports use the same active filters. Maker colors are editorial representatives; the logos come from recorded official sources. Financial figures retain their period, measurement basis and publication status. Interpretations are editorial; source links make the claims checkable.
- Artificial Analysis · model leaderboardIndependent
- Stanford · AI Index 2026Research
- Anthropic · Series HOfficial
- Reuters · reported IPO prospectusReported
- Reuters · compute commitmentsReported
- Anthropic · Opus 5.5Official
- Anthropic · Sonnet 5.5Official
- METR · task time horizonsResearch
- ARC-AGI-3 · launch baselineResearch
- ARC Prize · current leaderboardResearch
- Devin puts the agent in a workspaceOfficial
- MCP opens a tool-connection layerOfficial
- Devin becomes generally availableOfficial
- Copilot adds agent modeOfficial
- Claude Code enters research previewOfficial
- Codex delegates work to cloud sandboxesOfficial
- Copilot takes issues through to PRsOfficial
- Claude 4 and Claude Code GAOfficial
- Gemini CLI brings an open-source agentOfficial
- GPT-5-Codex specializes for agent codingOfficial
- Sonnet 4.5 pushes sustained codingOfficial
- Cursor studies long-running agentsResearch
- The Codex app coordinates parallel workOfficial
- ARC-AGI-3 tests unfamiliar interactive worldsResearch
- Cursor joins SpaceXOfficial
- Opus 5.5 advances the coding frontierOfficial
- Sonnet 5.5 raises the agent benchmarkOfficial
- DevDay expands shared agent workflowsOfficial
- Anthropic · company & governanceOfficial
- OpenAI · Codex CLI launchOfficial
- Cursor · background agents previewOfficial
- Cursor · 2.0 and ComposerOfficial
- Google · Antigravity launchOfficial
- METR · MirrorCode preliminary resultsResearch
- a16z · Consumer AI, edition 7Research
- SemiAnalysis · Subscription value testsResearch
- Anthropic · Haiku 5.5 release & pricingOfficial