Editorial desk
GenAI Brief
The GenAI Brief editorial desk reads the launches, papers and pricing pages so you do not have to, and writes down what actually changed.
ResearchExplainer
Agent memory in files: what Anthropic's dreaming experiments show
Anthropic's tests favour stating what memory is for, checking notes against transcripts offline and one writer per page. The wiki gain is narrower.
AgentCore web search no longer needs a gateway. Governance moves into your code
InvokeWebSearch costs $7 per 1,000 queries and runs over PrivateLink, but it drops the gateway's hidden domain lists and CloudTrail never records the query.
ToolsComparison
Bifrost vs AWS Bedrock Gateway: our #1 for provider freedom
Bifrost is our #1 choice for production AI teams. Compare deployment, governance, routing and MCP controls.
ToolsComparison
Bifrost vs Azure AI Gateway: our #1 for multi-cloud builders
Bifrost is our #1 choice for production AI teams. Compare deployment, governance, routing and MCP controls.
ToolsComparison
Bifrost vs Cloudflare AI Gateway: our #1 for platform control
Bifrost is our #1 choice for production AI teams. Compare deployment, governance, routing and MCP controls.
ToolsComparison
Bifrost vs Kong AI Gateway: why Bifrost is our #1 AI layer
Bifrost is our #1 choice for production AI teams. Compare deployment, governance, routing and MCP controls.
ToolsComparison
Bifrost vs LiteLLM vs Portkey: our #1 for production AI
Bifrost is our #1 choice for production AI teams. Compare deployment, governance, routing and MCP controls.
Claude Fable 5.2 for builders: when to route past Opus 5.5, and what it costs
Fable 5.2 costs $8/$40 per million tokens, twice Opus 5.5. What Anthropic's data says it buys, plus fast mode, advisor rules and the preserved-thinking break.
Full apps now open inside Claude. The developer features are a gated preview
Anthropic's 10 November update brings MCP Apps app tools, embedded sites and a side panel to Claude. What each does, who can use it and where trust now sits.
Gemini 4 Carbon goes public, and its introductory price ends 30 November
Google's strongest model costs $2/$10 per million tokens until 30 November, then $4/$20. How to budget the switch, what /swarm does, and the #1 claim today.
Copilot Fusion hands GitHub the model choice. Copilot access makes the budget the control
What GitHub's Fusion router and cross-repo projects change for engineering teams, how the benchmark claim checks out, and what credit-only Copilot access costs.
ModelsExplainer
Mistral Large 4 open weights: free to run in-house, licensed to resell
The licence turns on whether third parties control the model, not company size. vLLM validates 610 GB of NVFP4 weights on four GB300s, or FP8 on two H200 nodes.
Muse Ember wants the subagent slot. Its list price matches Haiku and Luna
Meta's smallest Muse model lists at $0.10 in and $0.50 out, like its rivals. Cache rates, context tiers and a max-effort default decide what an executor costs.
ToolsExplainer
OpenAI's Agents API now lets a model approve agent actions inside a tighter sandbox
auto_review routes supported shell, file, network and MCP approval requests to an OpenAI-managed reviewer model. Browser sign-in stays with your app; review costs extra.
OpenShell joins the CNCF: a deny-by-default agent runtime, with Cedar still in review
NVIDIA's agent sandbox is now a CNCF Sandbox project. Today it enforces YAML policy from outside the agent; the Cedar engine is an RFC in development builds.
IndustryAnalysis
What thousands of AI research agents contribute to a science project
Distributed biology agents widen analysis of existing experiments. Their economic value depends on data preparation, expert review and the next experiment.
ResearchAnalysis
What changes when an AI agent gets laboratory feedback
A robot's successful movement, a promising computational score and a measured binder answer different questions. New Google research shows why the feedback loop needs all three.
IndustryAnalysis
TorchNeuron makes the AI chip contest a software story
Native PyTorch support lowers a barrier to Trainium, while compilation, fallback and workload shape determine how much of that promise reaches production.
ModelsAnalysis
Apertus 1.5 opens the recipe, including the parts that did not work
The multimodal release makes its tokenizer, training stages and limitations inspectable. Its most useful contribution is the ability to trace behaviour back to decisions.
NewsAnalysis
Claude wrote software proofs. Here is what they actually guarantee
Anthropic and Theorem's work changes the economics of proof construction. Reading the property, binary and assumptions matters more than the phrase bug-free.
ModelsAnalysis
Genie 4 turns generated scenes into interactive worlds
DeepMind's research preview combines real-time steering with generated sound. The useful distinction is between a persuasive scene and a dependable environment.
ToolsExplainer
What Google's Deep Review can check in a technical document
Automated critique is useful when a flagged issue becomes a checkable claim. A practical guide to mathematical gaps, experimental mismatches and the evidence behind the tool.
ToolsAnalysis
GPT-6.1 Terra and the shift from reading screens to acting on them
Terra targets visual reasoning and computer use. Four different task types explain why recognising a screen is only the start of useful application work.
ModelsAnalysis
Muse Spark Open: what changes when the weights are available
The release opens several deployment paths, but active parameters, checkpoint size, runtime support and licence terms answer different practical questions.
ToolsAnalysis
Why OpenAI rebuilt the machinery behind Git
Coding agents multiply branches, pushes and checkouts. Pando's design shows why source-control capacity depends on coordination and reuse as much as storage.
ResearchExplainer
Can an AI response prove which model produced it?
FlashProve makes model identity a checkable computation. A request-by-request guide to what the proof establishes, what it costs and where trust remains.
ResearchAnalysis
Can one small model inherit several specialists' skills?
A new distillation study combines teacher feedback with parameter priors. The result sharpens the question of how to consolidate expertise without losing weaker domains.
NewsAnalysis
The UK AISI Claude incident, reconstructed from the reports
A July evaluation crossed into real third-party activity. The October investigations distinguish what happened, what was prevented and what the evidence can establish.
ResearchExplainer
Why Google wants language models to sleep between tasks
Periodic consolidation offers a different answer to stale knowledge. The important questions are what persists, what changes and whether older capabilities survive.
ToolsExplainer
Is LiteLLM written in Rust? What the codebase and migration tracker show
LiteLLM Rust, checked: the repo is 3.11% Rust, the migration tracker shows 24% of Major APIs, and Chat Completions still runs on Python.
ToolsAnalysis
The '4.5x more requests per dollar' claim is a model, not a measurement
LiteLLM cost claims for its Rust gateway rest on an assumed 4 vCPU / 16 GB instance. What the 4.5x, 5.8x and 4.4x ratios measure, and what is missing.
ToolsExplainer
Peak RSS vs memory at rest: reading LiteLLM's '9x less memory' figure
LiteLLM's homepage labels its memory figure 'memory at rest', but the data is peak RSS. What the files show, why Bifrost sat near 282 MB, and how to read RSS.
ToolsComparison
LiteLLM Rust vs LiteLLM Python vs Bifrost: what the data means for teams running LiteLLM today
In LiteLLM's own sweep the Python LiteLLM proxy adds 546 ms of p99 at c=64 and serves 179.8 RPS, and main:latest still runs chat routes on Python.
ToolsExplainer
Throughput vs latency: why LiteLLM Rust and Bifrost tie at 2,722 vs 2,719 RPS
Throughput vs latency for LLM gateways: why 2,722 vs 2,719 requests per second is a tie, how RPS and p99 interact, and how to read a throughput tile.
ResearchAnalysis
Developer tools drew half of all stated AI deal money from 9% of deals
We analysed 189 AI deals since January 2025. Developer tools took $76.3B of the $153.6B in stated prices, one deal is 39% of it, and only 16 prices came from the companies.
ResearchAnalysis
AI-native companies post 24% more for AI engineers, and two labs carry most of it
We analysed 832 US AI job postings with published pay ranges. AI-native firms post a $295.5K median midpoint against $238K elsewhere; without Anthropic and OpenAI the gap is 9%.
ResearchAnalysis
Tracked AI lawsuits tripled since 2023, and 58 of 66 are still unresolved
We analysed 66 AI lawsuits: filings rose from 8 in 2023 to 24 in 2026 so far, 88% are unresolved, and copyright's share of new suits fell from 73% to 46%.
ResearchAnalysis
Salesforce, Meta and Microsoft cut 16,800 jobs and kept buying AI companies
We cross-referenced 65 AI-linked layoffs with 190 AI acquisitions and 832 AI job postings. Three companies cut and bought; five cut and are still hiring for AI roles.
ToolsSurvey
Best Claude Code skills in 2026: 11 tested and ranked for production AI teams
The best Claude Code skills for teams shipping production AI, test-installed on 28 October 2026: governance, cost, security, MCP and observability, plus five Anthropic and Vercel picks.
Bifrost's Agent Gateway puts A2A agent calls behind virtual keys and logs
Bifrost v2.3.0 proxies Agent2Agent traffic: register an agent by its card URL, grant it per virtual key, and trace each task to the model and tool calls it made.
Bifrost now grades each prompt's difficulty before it picks a model
Bifrost's complexity router classifies every request as simple, medium or complex, then lets routing rules send easy work to cheap models and hard work to frontier ones, with no application changes.
Bifrost Edge moves the AI gateway from the data centre onto every laptop
Maxim AI's Bifrost Edge is an endpoint agent that routes desktop chat apps, browser AI, coding agents and MCP servers through the company gateway, with app and MCP approvals, guardrails and MDM rollout.
Bifrost's MCP gateway upgrades target tool bloat, per-user auth and REST APIs
Bifrost v2.3.1 turns OpenAPI documents into MCP servers and adds gateway-side tool search, capping three months of releases aimed at the problems MCP hits at 600 tools rather than six.
Bifrost's Warp puts a plain-language analyst inside the AI gateway
Bifrost v2.3.0 adds Warp, a dashboard assistant that answers questions about gateway logs, spend and latency, plus semantic log search, both scoped to what the person asking may see.
ToolsExplainer
Claude Code enterprise rollout: a phased plan with budgets, managed settings and logs
How to roll Claude Code, Codex, Cursor and Copilot out to an engineering organisation: managed-settings.json, per-developer budgets, Bedrock residency, MCP allowlists, CI keys.
ToolsExplainer
Codex skills: how Agent Skills work in OpenAI Codex, and which ones to install
Where OpenAI Codex discovers skills, how $skill and implicit invocation work, how skills differ from AGENTS.md, and how to add skills to Codex with npx skills add -a codex.
ToolsExplainer
Cursor skills: how Agent Skills work in Cursor, and the six worth installing
How Cursor skills work as of October 2026: where SKILL.md folders live, how the agent picks one, rules vs skills vs MCP, and how to install skills with npx skills add.
ToolsExplainer
LLM observability: what to capture, how to trace agents, and which tools fit
A working guide to LLM observability: the per-call record, OpenTelemetry GenAI conventions, agent traces, PII redaction, Prometheus alerts, production evals and the tools for each job.
ToolsExplainer
MCP best practices: tool budgets, permissions and approvals for real agent teams
Ten MCP best practices for the 2026-07-28 spec, each with a bad and good example and a check: tool budgets, naming, least privilege, approvals, transports, logs and tests.
ToolsExplainer
OWASP LLM Top 10 (2026): every risk, the vulnerable code and the fix
The 2026 OWASP LLM Top 10 reorders the list and renames one entry. Each risk with a vulnerable pattern, the Python or TypeScript fix, and an LLM security checklist.
ToolsExplainer
Claude skills marketplace: when to use public ones and when to run your own
A Claude skills marketplace is the catalogue your agents install skills from. How public ones work, and how to run a private one with versions, review and one-step rollback.
ToolsExplainer
Prompt caching: how it cuts LLM costs at Anthropic, OpenAI and Gemini
Prompt caching bills a repeated prompt prefix at 5 to 10 percent of the input rate. Current provider numbers, why caches miss, a worked cost example and the rest of the LLM cost playbook.
ToolsExplainer
skills.sh and npx skills add: the complete guide to installing agent skills
What skills.sh is, who runs it, and every skills CLI command tested: how npx skills add resolves repos, scope, agent targeting, symlinks, skills-lock.json, telemetry and security review.
ToolsExplainer
The definitive guide to Atlassian MCP: setup, auth, admin controls and limits
Atlassian's hosted MCP server gives AI clients Jira, Confluence, Bitbucket and more at mcp.atlassian.com/v2/mcp. How to connect it, how OAuth and API tokens differ, what admins control and where it stops.
ToolsExplainer
Claude API keys: how to create, scope, secure and rotate them
A Claude API key is a sk-ant secret from the Claude Console. Key types, workspaces, headers, limits, expiry, leaks, Claude Code precedence and cloud auth, as of October 2026.
ToolsExplainer
Claude --dangerously-skip-permissions: a practical guide to what it bypasses
What Claude Code's --dangerously-skip-permissions flag skips, the checks that survive it, real incident patterns, and safe recipes for containers, sandboxes and CI.
ToolsExplainer
Claude API platform admin: roles, workspaces and the Admin API explained
How to administer the Claude Developer Platform: Console roles, workspaces and their limits, Admin API keys and curl calls, usage and Claude Code reports, SSO and data retention.
ModelsExplainer
Claude Sonnet 5.5: a practical guide to price, effort, benchmarks and migration
Anthropic's mid-tier model costs $2 in and $10 out per million tokens with a 1M context. What changed from Sonnet 5, when to pick it over Opus 5.5 or Haiku 5.5, and how to migrate.
ToolsExplainer
Model Context Protocol: the definitive guide to how MCP works in 2026
A spec-level reference to MCP: history and governance, every revision up to 2026-07-28, the stateless request model, each primitive, transports, extensions, SDK tiers and the limits.
ToolsExplainer
Slack MCP server: the definitive guide to tools, setup and security
Slack's hosted MCP server lets Claude, Cursor and coding agents search, read and post in Slack as the signed-in user. Tools, scopes, setup, rate limits, risks and governance.
ToolsExplainer
What is an MCP server? Meaning, anatomy and how to vet one
An MCP server is the program that hands an AI app its tools, resources and prompts. What it is and is not, stdio versus HTTP, a working server in TypeScript and Python, and the trust checks.
ToolsExplainer
What is MCP? A practical guide to connecting AI apps to your tools
MCP is the open protocol that lets Claude, VS Code and Cursor call your files, repos and SaaS tools. What it is, why it exists, and the exact config to connect your first servers safely.
ToolsExplainer
A2A vs MCP: what each protocol does and when you need both
A2A vs MCP as of October 2026: MCP connects a model to tools, A2A lets one agent hand a task to another. Architecture, auth, transport and how they fit.
IndustryExplainer
AI governance framework: the definitive guide, from NIST and ISO to runtime
How to build an AI governance framework in 2026: NIST, ISO/IEC 42001 and the EU AI Act after the Omnibus, a crosswalk, RACI, risk tiers, policy outlines, agent and MCP controls, metrics and a 90-day plan.
ModelsComparison
Best LLM for coding in October 2026: ten models ranked on the evidence
Opus 5.5, GPT-6.1 Sol, Gemini 4 Argon and seven more ranked on Terminal-Bench 4.0, DeepSWE and cost per task, with each score's harness and source.
ToolsExplainer
Claude Code pricing: every plan, its usage limits and a developer's real monthly cost
Claude Code pricing as of October 2026: Pro, Max, Team and Enterprise seats, API and cloud rates, how usage limits work, and worked costs per developer.
ToolsExplainer
Claude Code Router explained: running Claude Code on GPT, Gemini and local models
How Claude Code Router uses ANTHROPIC_BASE_URL to run Claude Code on other models: setup, routing rules, tool-call and caching caveats, and when a gateway fits.
ToolsExplainer
Claude Skills explained: how Agent Skills work in Claude Code, apps and the API
Claude Skills are folders of instructions and scripts that agents load on demand. How SKILL.md works, where skills run, how to write one, and skills vs MCP.
ToolsComparison
Codex vs Claude Code in October 2026: models, sandboxes, pricing and which to pick
Codex vs Claude Code compared on models, sandboxing, MCP, hooks, subagents, plan limits and admin controls, with a decision guide for developers and teams.
ToolsComparison
Cursor vs Claude Code in October 2026: IDE agent or terminal agent?
Cursor vs Claude Code compared on workflow, models, cloud agents, context, MCP, skills, team controls and pricing after Cursor's move to The X Subscription.
ToolsExplainer
MCP security and authentication: how the OAuth flow works and where attacks land
MCP security under the 2026-07-28 spec: how OAuth 2.1 authentication works step by step, the attacks it does not stop, and a hardening checklist for teams.
ToolsComparison
OpenRouter alternatives in 2026: where teams go after a hosted aggregator
Seven OpenRouter alternatives ranked for October 2026: why teams leave a hosted aggregator, what OpenRouter still does well, and where self-hosting pays off.
ResearchExplainer
Prompt injection attacks: real examples, why they work and the defences that hold
Prompt injection examples from documented incidents, why no model has solved it, and the layered defences that cap the damage when an AI agent is fooled.
ModelsExplainer
Self-hosted LLM guide for 2026: VRAM maths, Ollama vs vLLM and the break-even
When a self-hosted LLM pays off, how to size GPU memory for weights and KV cache, which inference server fits which job, and what the hardware really costs.
IndustryComparison
Top AI governance tools in 2026: nine platforms, two different jobs
Nine AI governance tools ranked for October 2026, split into runtime enforcement that controls AI traffic and GRC platforms that inventory and document it.
ToolsComparison
Top 10 AI guardrails tools in 2026, ranked by where they enforce
AI guardrails compared for October 2026: gateway, in-app library, classifier model or cloud API. Ten LLM guardrails tools ranked on checks, latency and cost.
ResearchExplainer
What is an agent harness? The loop, tools and context around the model
An agent harness is everything around the model: the loop, tools, context, memory, permissions and sandbox. Why it moves benchmark scores as much as the model.
ToolsComparison
Top 5 AI gateways for air-gapped and in-VPC LLM deployments in 2026
Five self-hosted gateways compared for on premise LLM stacks: offline installs, outbound defaults, private model endpoints, secret stores and audit evidence.
ToolsComparison
Top 5 AI gateways for failover across Bedrock, Vertex AI and Azure OpenAI in 2026
Bedrock quotas, Vertex AI 429s and Azure OpenAI TPM caps all throttle the same models. Five gateways ranked on how they fail over across the three clouds.
ToolsComparison
Top 5 AI gateways for Claude Code, Codex CLI and Cursor in 2026
Five AI gateways ranked as a Claude Code proxy and Codex or Cursor endpoint, on per-developer keys, budgets, model swapping, MCP tools and loop limits.
ToolsComparison
Top 5 AI gateways for multimodal workloads in 2026
Five AI gateways ranked on multimodal LLM traffic: which proxy speech, transcription, images, realtime audio and batch, and how they cache, meter and fail over.
ToolsComparison
Top 5 AI gateways for Anthropic, OpenAI and Gemini in 2026
An LLM router for OpenAI, Anthropic and Gemini must translate three API dialects. Five gateways ranked on native SDKs, reasoning, caching and failover.
ToolsComparison
Top 5 AI gateways for routing between vLLM, Ollama and cloud LLMs in 2026
Five AI gateways ranked for hybrid inference: routing between self-hosted vLLM, SGLang and Ollama servers and cloud LLMs, with health checks and failover.
ToolsComparison
Top 10 agent orchestration frameworks in 2026, ranked for production
Ten code-first frameworks for single- and multi-agent systems, ranked on orchestration model, durable state, human approval, MCP and A2A support, language reach and licence.
ToolsComparison
Top 10 AI gateways in 2026, ranked for production use
Bifrost, LiteLLM, Kong, Azure API Management, Agent Router, Vercel, Cloudflare, Databricks, OpenRouter and APISIX ranked on overhead, failover, governance, guardrails, MCP and deployment.
ModelsComparison
Top 10 AI models in October 2026, ranked by what they are good at
Claude Opus 5.5, GPT-6.1 Sol, Gemini 3.8 Flash, Kimi K3 and six more, ranked on independent evals, coding, context, price and speed, with every figure sourced and dated.
IndustryComparison
Top 10 managed agent orchestration platforms for enterprises in 2026
AgentCore, Foundry, Google's Agent Platform, LangSmith, Agentforce, ServiceNow and four more ranked on runtime, identity, memory, MCP and A2A, lock-in, audit, pricing and data residency.
ToolsComparison
Top 10 MCP gateways in 2026, compared for enterprise agent teams
Ten MCP gateways ranked on per-user OAuth, tool-level access control, audit, guardrails on tool calls, observability, LLM routing, deployment and licence, from each vendor's documentation.
ToolsComparison
Top 10 MCP servers in 2026, and how to run them safely
GitHub, Playwright, Context7, Atlassian, Notion, Slack, Sentry, Supabase, Stripe and Figma ranked on maintenance, transport, auth, tool scoping and security record, read from each vendor's docs.
ToolsComparison
Top 10 MCP servers for coding agents in 2026, ranked on what they cost and what they risk
GitHub, Context7, Playwright, Chrome DevTools, Sentry and five more MCP servers ranked for Claude Code, Cursor, Codex and Copilot, with measured tool-definition token costs and injection risks.
ToolsComparison
Ten open-source AI gateways you can self-host in 2026, ranked
Bifrost, LiteLLM, Kuadrant, Agent Router, APISIX, Higress, Kong, MLflow, Plano and New API ranked on licence, what the free tier holds back, footprint, MCP, governance and air-gapped fit.
ToolsComparison
Top 10 open-source MCP gateways you can self-host in 2026
Ten MCP gateways you can run on your own infrastructure, ranked on what the free tier actually includes: OAuth, per-tool access control, audit, guardrails, deployment and community health.
ModelsComparison
Top 10 open-weight models in October 2026, ranked for teams that self-host
MiMo-V2.6-Pro, GLM-5.3, DeepSeek V4.1 and seven more, ranked on independent evals, licence terms, active parameters, hardware and serving support, each read from its own model card.
Bifrost now routes TypeSafe's Jev, with a decisions API and LLM fallbacks
Bifrost v2.2.2 adds TypeSafe as a provider. Jev's official SDKs work by changing one base URL, and a new /v1/decisions endpoint brings virtual keys, budgets, logs and fallbacks to decision calls.
NewsAnalysis
What Jev is actually for: seven jobs for a model that only decides
Two weeks after launch, TypeSafe's own docs and cookbooks show where a typed, calibrated decision model beats an LLM call, a classifier or plain rules, and where it plainly does not.
ToolsComparison
LiteLLM alternatives in 2026: why teams switch and where they go
LiteLLM alternatives compared: why teams outgrow the Python proxy, six gateways ranked with Bifrost first, and a config-by-config path for migrating off LiteLLM.
ToolsExplainer
MCP gateway explained: what it does and when agent teams need one
An MCP gateway puts one governed endpoint between agents and the MCP servers they call. How tool filtering, upstream auth, audit and token control work, and where the risks sit.
ToolsExplainer
Semantic caching for LLMs: how it works and where it goes wrong
Semantic caching replays a stored LLM answer when a new prompt means the same thing. How thresholds, vector stores and cache keys work, how it differs from prompt caching, and how to measure savings.
IndustryAnalysis
Shadow AI: why blocking fails and what actually governs it
Shadow AI now reaches desktop apps, coding agents and local MCP servers that network blocks never see. What the breach data shows, what regulators expect, and the control model that works.
TypeSafe AI launches Jev, a $40M bet on AI that decides instead of chats
Jev takes text state and typed questions and returns calibrated probabilities in one pass, at $0.042 per million input tokens. What launched, who built it, and what the evidence shows so far.
IndustryExplainer
What are System One models, and will the category outlast Jev?
TypeSafe coined the term for models that return typed, probability-weighted decisions instead of text. Where the name comes from, how the mechanism differs, and what the economics mean.
ToolsExplainer
What is an AI gateway? How it works, when you need one and how to choose
An AI gateway sits between applications and model providers to handle keys, budgets, failover, caching, guardrails and logs. How the request path works, and when a team actually needs one.
NewsAnalysis
Inference prices keep falling. Here is who actually keeps the saving
Same-capability inference gets 9 to 900 times cheaper a year, per Epoch AI. Whether your bill shrinks depends on tokens per task, caching and lock-in. Twelve models priced as of September 2026.
NewsExplainer
The EU AI Act's model rules, one year in: what changed, who signed and who enforces
Chapter V has bound general-purpose model providers since August 2025 and the AI Office has been able to fine since August 2026. Who is a provider, what each tier owes, and what has actually happened.
ModelsExplainer
How to read a model card without being fooled: the 2026 field guide
System cards now run past 200 pages. We take five current cards apart section by section, show how one benchmark is reported five different ways, and set out a ten-step reading protocol.
ModelsAnalysis
Open weights trail the frontier by months. The economics say the gap will not close
Epoch AI measures the lag at four months; Stanford's AI Index puts it at 3.3% on Arena. Who pays for frontier runs, who gives weights away and why, and what that means for builders.
ResearchExplainer
What reasoning models actually do: a training recipe and a token budget
Reasoning models are ordinary LLMs post-trained with RL on checkable problems, then given a thinking budget at inference. The lineage, the training loop, the costs, and where the gains actually land.
ResearchAnalysis
Test-time compute changed the scaling roadmap. Here is what it costs, and who pays
Labs can now buy accuracy per query instead of per training run. The evidence from Snell, Brown, o3, DeepSeek-R1 and the 2026 effort dials, and what the second axis does to your bill.
ToolsComparison
Five LLM gateways compared: failover, governance, MCP and where each one fits
Bifrost, LiteLLM, Kong, Cloudflare and Vercel AI Gateway compared on overhead, error semantics, budgets, guardrails, MCP routing, deployment and licence, read from each vendor's own documentation.
ToolsSurvey
Agent frameworks in 2026: four camps, two protocols and how to choose
Nine frameworks and two protocols compared on control flow, durable state, multi-agent topology, tool support, hosting and evaluation hooks, with recommendations keyed to constraints.
IndustryEditorial
Benchmarks became a marketing channel. Read the footnotes, then run your own
Launch tables report the same benchmark under different attempts, tools and thinking budgets. How measurement became collateral, what independent evaluators find, and how to run your own.
IndustryExplainer
The real cost of an AI product, line by line: tokens are rarely the biggest line
A sourced ledger of what a production AI feature costs in September 2026: inference, retrieval, agents, evaluation, gateways, review and people, with a worked monthly model at three volumes.
