Insights

Our Agents Never Picked Canva. 10 Learnings on Agent Pricing, Getting Picked, and Security From This Week’s “The Agents” With Ulrik Lehrskov-Schmidt

Amelia and I both canceled Canva this week. I’d been a customer for about 12 years. We weren’t angry at it, and nobody sat down and decided to replace it. Our agents just never had a use for it. That happened the same week Canva restructured 22% of the company. Melanie Perkins said in her... Continue Reading

Higgsfield Went From $1B to $1.35B in Run Rate in One Month. Now It’s Aiming for $2B by Year-End.

Disclosure: Higgsfield is a SaaStr Fund portfolio company. We invested early, and we’ve been customers since the early days. Higgsfield crossed $1 billion in annualized revenue run rate in September. Alex Mashrabov just sent me the October number: $1.35 billion. That’s $350 million of new run rate in about a month. Most companies at that... Continue Reading

BotPsy: On-Device Personal Assistant with a Sub-Billion-Parameter Agentic Language Model

Token Pooling-Aware Training for Multi-Vector Embedding Models

Gemini Decision 1 now available on AI Gateway

is now available on AI Gateway. The multimodal decision model works with text, images, audio, video, and PDF documents, returning Boolean probabilities, choices, and scores. Use it to route requests, check policies, filter agent tools, or rank results without generating output tokens.Gemini Decision 1 from Google Use with the AI SDK, TypeSafe-compatible API, or Decisions API.google/gemini-decision-1 PDF, audio, and video inputs are supported through AI SDK 7. The TypeSafe-compatible and Decisions APIs support text and image inputs. AI Gateway provides a unified way to access models and track usage and cost. Decision calls appear in and count toward alongside other model requests.logsbudgets See the for Choice, Score, and Boolean questions. To pick a decision model, browse the or read our .decision documentationavailable modelsdecision model guides Read more

Anthropic’s IPO Puts ~$216 Billion in Employee Hands. Another $160 Billion Is Reserved for the Next Hires.

Anthropic's employees hold ~$216B of equity at the IPO price. Another $160B is reserved for future hires. 5,300 employees. ~$41M each on average. Every B2B founder now competes for talent against liquid ANT stock.https://t.co/lMJbuEJzEQ pic.twitter.com/T6hdBolVeU — Jason ✨👾SaaStr.Ai✨ Lemkin (@jasonlk) November 10, 2026 Anthropic’s amended S-1 is out, and most of the coverage is about... Continue Reading

AI Startups May Have Fewer Sales Reps. But Not as a Percent of the Team. Has The Playbook Really Changed All That Much?

The common view for the past 12-18 months has been that AI-native companies reach $100M+ ARR with very small teams, that agents handle much of the selling, and that the traditional B2B sales org is shrinking away. Part of that holds up. AI-forward companies do run smaller go-to-market teams in absolute terms. ICONIQ’s State of... Continue Reading

Pay for Hugging Face Hub with your Google Cloud Account

Decision Models in Hugging Face

MiniMax H3 Image and Highspeed now available on AI Gateway

and are now available on AI Gateway. Generate images from text or edit existing images with one API key and no separate MiniMax account.MiniMax H3 ImageMiniMax H3 Image Highspeed Both models support custom dimensions and aspect ratios, including 2K and 4K outputs. Use a text prompt to create a product visual or illustration, then pass the image back with instructions to change colors, subjects, or composition. The Highspeed variant offers faster generation with the same generation and editing options. Use or as the model name:minimax/minimax-h3-imageminimax/minimax-h3-image-highspeed To edit the generated image, pass it back with an instruction: Both models also support the OpenAI-compatible and routes. See the for details./v1/images/generations/v1/images/editsimage generation guide Try MiniMax H3 Image in the .model playground AI Gateway provides a unified API for calling models, tracking usage and cost, and managing access with one API key. View requests and spend in your dashboard. AI Gateway reflects provider pricing with no markup and no platform fee on inference. See for details.AI Gateway pricing You can view available on AI Gateway.all image models Read more

Gemini 4 Carbon now available on AI Gateway

is now available on AI Gateway. It is available at 50% off through November 30, 2026:Gemini 4 Carbon from Google Gemini 4 Carbon is built for complex software engineering, sustained agentic work, and cybersecurity defense. It has a 1M-token context window for working with large codebases and long inputs. On engineering tasks, Carbon carries demanding work through multiple stages, with improvements in coding, full-stack web development, and scientific reasoning. Its cybersecurity capabilities span finding vulnerabilities, validating their risk, and developing patches. more reliable execution of multi-step plans, fewer unsupported claims, and better collaboration when requirements are ambiguous, helping agents follow through while respecting project constraints.Google also reports Use as the model name:google/gemini-4-carbon To use it in Claude Code, Codex, Cursor, and more, install the latest Vercel CLI and run setup: Then select in the agent. See the for details.google/gemini-4-carboncoding agents guide Try Gemini 4 Carbon in the .model playground AI Gateway provides a unified API for calling models, tracking usage and cost, and configuring retries, failover, and performance optimizations for higher-than-provider uptime. It includes built-in , , , , and more.custom reportingZero Data Retention supportbudgets for API keysrouting rules AI Gateway reflects provider pricing with no markup and does not charge a platform fee on inference, including on (BYOK) requests.Bring Your Own Key You can view available on AI Gateway.all language models Read more

Speed Insights and Web Analytics now support custom environments

and now fully support .Speed InsightsWeb Analyticscustom environments Previously, data captured on deployments to a custom environment was attributed to the Preview environment. That data is now attributed to the custom environment it came from, so you can track performance and traffic for staging, QA, or canary environments separately from Production and Preview. You can view this data in several places: This change only affects how data is attributed. No new data points are collected, and no configuration changes are needed. Custom environments are available on Pro and Enterprise plans. Learn more in the .environments documentation Read more use the Environment dropdown in Speed Insights and Web Analytics to pick a custom environmentDashboard: query it with the commandCLI:vc metrics filter and break down by environmentQuery builder: query it programmatically (Web Analytics only)Aggregation API:

Vercel Sandbox snapshot creation is now up to 10x faster

now creates snapshots up to 10x faster. Median time to snapshot is under 0.4 seconds, down from 4.4 seconds. Snapshot restore speed has also improved by 50% from 500ms to 250ms at p50.Vercel Sandbox Snapshots are now able to write and restore by re-using frames within the Vercel Hive Snapshot (VHS) format. VHS now indexes data in a more efficient way which means that more read and write operations can be done in parallel. Snapshots can now restore more of your Sandbox base image in parallel with user data, leading to overall faster operations. Persistent sandboxes, the default, snapshot automatically when a session stops and resume from the latest snapshot. This applies to all new snapshots, with nothing to configure. Existing snapshots keep the previous format. Learn more in the .snapshots documentation Read more

Nemotron-Personas-Finance-Japan: A Synthetic Dataset Capturing Japan's Demographics and Financial Context

The New Normal in Revenue Per Employee: $237K Is the OldMedian, $655K Is the New Top Tier, and Six B2B Leaders Now Clear $1.5M

For 15 years, B2B ran on one planning ratio that almost no one questioned: Every $200K to $300K of new revenue needed another person. Boards approved hiring plans against it, CFOs modeled against it, and founders raised money to fund it. That ratio still describes the median older B2B / SaaS company. But it no... Continue Reading

Your Experts Have Partners. There's HOPE for Keeping Them Together

Decision models take 26% of requests; Anthropic now wins new teams 2:1

AI Gateway Production Index — November 2026 Every month, routes more than 100 trillion tokens between production applications and AI labs. That traffic gives us a view of what AI usage looks like in today’s enterprise and startup landscape across language, decision, voice, image, and video models. We publish it here monthly. See the Production Index reports from , , and .AI GatewayAugustSeptemberOctober Production AI is splitting into two jobs: generation and judgment. Until recently, when software needed to make a quick call, a team could use either a fixed rule that was cheap but rigid, or a language model that was smart but slow and expensive. Decision models are now a third option. They classify, score, or pick between choices instead of generating text, and cost about a seventieth as much per request as a language model. They arrived on AI Gateway in mid-September, and there are now 15 decision models available from eight different vendors. In October, nearly a third of paid teams on AI Gateway used a decision model. About 11% of paid teams used a decision model without sending a single request to a text-generating model. Decision models cost far less per token and give very short answers, so even though they received about one in four gateway requests in October, they accounted for just 0.53% of total spend. Almost all of that request volume went to Jev from TypeSafe AI, which alone received nearly four times Anthropic’s total requests. For every $1 teams spent on Jev, they spent $64 on Anthropic models. Decision models are most common among the teams that run the most AI. Just over one in five teams running fewer than 10 million tokens used a decision model in October. At 100 million to 1 billion tokens, nearly half used one. Above 10 billion, seven in ten did. This is partly because big teams use more of everything. The top 1% of teams by spend use a median of 29 different models, and the bottom tenth use only one. It’s also volume. The more requests a team sends, the more it has to gain from a model that makes simple decisions faster and far more cost-effectively. About one in nine paid teams used a voice model in October, up from one in 24 in July, an increase of 161%. New teams are using voice, and existing teams are adding it. Among teams that paid in both July and October, the share using voice went from 5% to 14%. Teams that use voice are also using more of it. The share of teams sending at least 100 voice requests a month grew fivefold, and six in ten of September’s voice teams returned in October, up from fewer than half in July. A spoken conversation with an AI app takes three steps: Most voice teams use just one or two of these pieces. Use of speech-to-text models has nearly tripled since July, and text-to-speech has more than tripled. Some models do all three steps on their own. In July, about a quarter of voice teams used one of these all-in-one models, and one in six assembled the pieces themselves. By October, those numbers had swapped. For most teams, voice is a new way for users to talk to an app or agent they already have. Teams’ preferred language models stay in the middle, and they can add voice at one end or both. Last December, 70 of every 100 paid teams on AI Gateway used an OpenAI model and 48 used an Anthropic model. By October, the two labs were nearly level, with OpenAI at 55 and Anthropic at 54. Both labs now reach more of the teams that have been on the gateway since December, but Anthropic’s reach grew more. The share of these teams using Anthropic rose from about half to two-thirds, while OpenAI’s rose from about three in four to four in five. Most of Anthropic’s gain came from new teams. In October, it was the first-choice lab for nearly twice as many of them as OpenAI. But among the biggest teams, both labs already have nearly everyone. More than nine in ten of the top 1% by spend use both Anthropic and OpenAI. With these teams, the labs now compete for how much of the work each one gets. On that measure, Anthropic still leads, but OpenAI has gained ground. OpenAI now takes nearly 15 cents of every dollar spent on the gateway, up from about 9 in December, while Anthropic’s halved to about 32 cents over the same period. Teams are moving to new models almost as fast as labs release them. More than half of October’s language-model requests went to models less than 60 days old. Before September, that share had never passed three in ten. Within each model family, a new version arrives every 61 days at the median, and a typical model peaks within a month of launch. Four weeks later, three in four of the teams using it at its peak aren’t using it anymore, whether or not a successor has shipped in the family. Most of those teams are still on AI Gateway, using other models. Teams switch models frequently, usually to something newer, and often to another lab. In October, about half of teams running at least 10 million tokens a month changed their most-used model. Four in five of those moved to a more recent release, and six in ten moved to a different lab. Older models stay in use, often alongside a newer primary model. Seven in ten teams running at least 10 million tokens a month still used a model at least ten months old in October, but for the typical team it ran less than 3% of its tokens. AI models now have a half-life of about six weeks. That means that, to the teams using them, labs are only as good as their latest model and have to re-earn their business with every new one. This report uses anonymized, aggregate traffic routed through Vercel AI Gateway through October 2026. A few notes on measurement: Read more about our , and .methodologyexplore or download the public data Read more November 2026 Summary About this data In six weeks, decision models grew to gateway requests, while accounting for of every dollar spent. of all paid teams used one in October. Among teams processing over 10 billion tokens a month, used a decision model.one in fourhalf a centClose to a thirdseven in ten About of paid teams used a voice model in October, up from in July.11%4% OpenAI and Anthropic now reach about the of paid teams, and new teams in October picked Anthropic . OpenAI’s share of gateway spend, meanwhile, is up by since December.same sharenearly two to onetwo-thirds of language-model requests went to models less than two months old, up from less than three in ten. Within one month of a typical model’s peak usage, of the teams using it have moved on.More than halfthree in four A voice model turns the user’s speech into text A language model writes the reply Another voice model reads the reply back to the user Opus 5.5 was the top model by spend in its first full month. It took about one in seven dollars across all teams. Four of the six most widely used models on AI Gateway are Anthropic’s: Sonnet 5.5 (reaching 26% of paid teams), Haiku 5.5 (22%), Haiku 4.5 (20%), and Opus 5.5 (18%). Jev (28%) and GPT-6 Luna (19%) are the other two. Image and video both had their biggest month yet on AI Gateway. More teams than ever are putting spend on them, and the largest teams are adopting them the fastest. About a third of the top 1% of teams by spend now use an image model, up from a quarter in July. One in six use a video model, up from one in ten. OpenAI’s newest image model, GPT-Image-2.5 Sunburst, took the lead in image spend in its first full month, followed by its Flare and GPT-Image-2 models. ByteDance’s Seedance 2.5 led video spend for a third month running, ahead of Google’s Veo 3.1 and Kuaishou’s Kling 3.0. Spend is estimated at published list prices; actual bills may differ. Paid teams are teams with at least one paid request in the month. Team size is measured by a team’s total tokens that month. Decision models are the models classified as decision models in the AI Gateway model list. Each model version is counted separately. A model’s age is measured from that version’s first day on AI Gateway, and gateway history begins October 20, 2025. Switching rates compare paid teams with at least 10 million tokens in both September and October. A team’s most-used model is measured by tokens. Voice figures begin June 22, 2026. Teams that “assembled the pieces” used speech-to-text, a language model, and text-to-speech on the same day. A new team’s first-choice lab is the lab it sent the most language-model requests to in its first month. All figures use the most recent data and protocol available; prior months may be revised as methodology is updated. Decision models account for 26% of all requests Voice adoption is up 2.6x since July Anthropic wins new teams two to one, and OpenAI’s share of spend is up two-thirds Most AI traffic now runs on models less than two months old Also in October’s data

GLM 5.3 Flash is now available in Token Pass

now includes , alongside Kimi K3, GLM 5.3, and DeepSeek V4.1 Flash (with more models to come). GLM 5.3 Flash is Z.ai's native multimodal coding model with a 1M-token context window.Token PassGLM 5.3 Flash Token Pass is a $30/month subscription on Pro teams that grants $80 in monthly credit and runs through . Use it with any coding agent such as , , , etc. or build your own with the AI SDK, with Zero Data Retention enabled by default.AI GatewayfxPiClineCommand Code Add GLM 5.3 Flash from the in your team's AI Gateway dashboard, or read the for setup.Token Pass pageToken Pass documentation Read more

Workflow SDK now writes over WebSockets on Vercel

The now sends event writes and stream writes to over a WebSocket by default, replacing one HTTP request per write, or per write group for streams. No code change is needed beyond upgrading to or later.Workflow SDKVercel Workflowsworkflow@5.2.0 Once you deploy the upgrade, new runs use it. In-flight runs finish on the deployment they started on. Existing streaming code works as is. A step that writes a burst of chunks, like AI tokens, sends them over the WebSocket once it opens: In a replaying an AI agent's token stream, v5.2.0, which writes streams over a WebSocket by default, delivered the full response 4.2× faster than v4.8.12, with no chunk delayed over a second.benchmark Every step start and completion is an event write, now sent over the WebSocket, so workflows with many steps make fewer HTTP requests. The added efficiency also improves performance and reduces costs. The queue handler opens the connection for every delivery. Read the to get started. If you're on 4.x, start with the , since v5 is a major release.Vercel World documentationupgrade guide Read more Streams keep pace with AI output Fewer HTTP requests for event writes Get started

ToyBlocks: A controlled world for robot learning and evaluation

Egyptian and Nigerian English join the Open ASR Leaderboard

Should You Visit More Of Your Prospects In Person? Almost Certainly

I remember the first customer I lost due to not showing up in person.  They were a Fortune 50 customer.  We did a CSAT survey, and every user loved us.  The implementation was flawless.  There were zero issues.  And … we lost them at renewal. One thing that always works: meeting more customers IRL Be... Continue Reading

Customers Love to Talk to the CEO. Do More of That.

Customers love, love, love to hear from the CEO This alone is one reason why start-ups have an edge Take advantage: – Go talk to 10+ of them a week– Do a user conference, it's OK to start small– Meet every larger customer in person– If there are bumps, just email them… — Jason ✨👾SaaStr.Ai✨... Continue Reading

Dear SaaStr: How Do You Respond to a Prospect That Sees You As An “Acquisition Risk”?

Dear SaaStr: How Do You Respond to a Prospect That Sees You As An “Acquisition Risk”? “Acquisition risk” is a subset of a whole category of “vendor risk”. Until you IPO, or are the de facto leader over many years, Big Cos are always going to look at “vendor risk.” My #1 answer in Vendor... Continue Reading