CometAPI vs Orthogonal: Same Gateway Label, Two Different Jobs

Hokai comparison of the two billing meters, from both vendors' pricing pages as read on 7 November 2026.

CometAPI and Orthogonal both call themselves a unified AI gateway, and as of November 2026 that is still where the resemblance ends. One sells model tokens at a fixed discount. The other sells tool calls to agents. Almost nobody writes them up side by side, because almost nobody treats them as the same category.

What changed since 13 August 2026

We re-read both vendors on 7 November 2026 and kept the verdict. The split holds. Several of the details we printed in August did not hold, so here is what moved:

Two claims from the August version are cut because we could not re-confirm them: CometAPI's "unlimited requests per minute" on Enterprise, and the line that neither company had a SOC 2 or GDPR attestation. We did not open either vendor's trust page this time, so we make no compliance claim in either direction.

Who is this comparison actually for?

If you build with AI models directly, you have probably heard of CometAPI. If you build AI agents that take actions in the world, you may have heard of Orthogonal instead. Rarely both. That split is the whole story here.

A two-person team shipping an agent product in 2026 does not pick one gateway and get everything. It picks a model router for the reasoning layer and, separately, a tool-access layer for anything the agent needs to touch outside the model call.

Pick CometAPI when the job is sending calls to a GPT, Gemini, Kimi, image or video model at a discount. Pick Orthogonal when the job is letting an agent enrich a lead, scrape a page or look up a company without you writing a separate client for each provider.

What is actually in each catalog?

CometAPI's catalog is models. Point an OpenAI-compatible client at its endpoint, swap the model name, and reach the 500+ models its site advertises, plus image, video and audio generation, according to CometAPI's pricing page. There is no agent-tool layer. It does not fetch a webpage or send an email. It returns a completion, an image or a clip.

Orthogonal's catalog is the mirror image. Its home page lists 60+ providers behind one key, sorted into five groups: web search (12 providers), scraping (15), enrichment (27), agent actions (6) and market and social data (16).

Apollo.io sits inside it as a lead-enrichment partner. Tavily sits inside it as a search and crawl provider, and you can also buy Tavily directly. None of it is a foundation model. An agent calls Orthogonal to look something up or take an action, then still needs a model to decide what to do with the result.

Access runs through three surfaces, per Orthogonal's docs:

• a hosted MCP server at mcp.orthogonal.com,
• a REST API at api.orthogonal.com with search and run endpoints,
• a CLI for coding agents such as Claude Code and Cursor.

That is deliberate. Orthogonal assumes the caller is an agent, not a person filling in a dashboard form. The agent searches the catalog in plain language, reads the price of an endpoint, and then runs it.

How do the two bills compare?

The number everyone expects is the model discount. CometAPI says it prices official models at 0.8 times the provider's own rate, with no monthly fee and no minimum spend. On 7 November 2026 its table showed:

That is a real, flat 20% across every row we checked. Credits do not expire, and the page says new accounts get free trial credits with no card. Enterprise buyers negotiate custom pricing, a dedicated account manager and custom service terms by email, so there is no published enterprise price.

Orthogonal does not compete on that number, because it is not selling model tokens. It bills per call from a pre-paid balance, with no subscription. Its docs group endpoints into three bands:

• LinkUp AI search$0.004
• Olostep scraping$0.005
• Hunter email search$0.01
• Apollo person enrichment$0.03 Leader

Read from Orthogonal's pricing documentation.

New accounts get $5 in free credits, no card needed. Every response also carries the cost of that call, so spend is visible per action. Agents can pay with stablecoins over the x402 or MPP protocols instead of buying credits.

There is still no flat rate card, and that is a real weakness. Budgeting an agent's monthly cost means running it and watching the meter. Comparing the two bills is comparing a per-token price to a per-call price for services that do not overlap. The honest question is which meter applies to the thing you are building. For model-only cost across vendors, our cheapest LLM API guide is the better starting point.

Where does CometAPI win?

Pick CometAPI if you are building a chat product, a coding assistant or anything that routes between several LLMs and keeps the code identical across providers. Its real competitive set is the model routers, not Orthogonal:

• OpenRouter, the best-known router, covered in our analysis of the Stripe and OpenRouter deal;
• LiteLLM, which you run yourself;
• APIMODELS, another hosted gateway;
• Together AI and Replicate, which host open models directly.

We did not re-price these in this pass, so each one's own page is the place to check current rates. What we can say from CometAPI's own page is the shape of its pitch: a flat 20% discount, one invoice for every provider, and image, video and audio models behind the same key. Its FAQ argues it beats OpenRouter on cost because OpenRouter charges a fee on credit purchases. That is a vendor claim about a competitor, and we have not tested it.

A four-person startup comparing Claude Opus 5 with Gemini 3.6 Flash on the same prompt, without a separate contract for each vendor, is who this is built for. So is a solo developer prototyping a video feature who does not want a new vendor relationship just to test Flux 3. If your test set also includes DeepSeek V4 Flash, a router means one key instead of three.

Where does Orthogonal win?

Pick Orthogonal if your agent needs to act, not just answer. A founder building a sales-outreach agent that pulls a contact's title from Apollo.io, checks the company, and returns a summary has no reason to write three separate API clients. One key and one MCP address cover them, and the response states the exact cost of each step.

The MCP-native setup is the other real difference. Orthogonal's docs show a single JSON block that adds its connector to Claude, Cursor or any MCP client, followed by an account sign-in. CometAPI has no MCP layer in anything we read. It is a plain HTTP endpoint built for a server calling it, not for an agent discovering it.

The company side is small. Orthogonal's home page announces a $4.3 million round led by Pantera Capital. Its Y Combinator profile lists a Winter 2026 company founded in 2025, with a team size of two. That track record rules nothing out for a startup, but a regulated enterprise buyer will want a longer history and a trust page before signing. CometAPI publishes enterprise terms only on request.

Is Orthogonal turning into a model gateway?

The strongest objection to calling these unrelated: Orthogonal's YC listing still says it is starting with "Search, Scraping, Datasets, and AI Model APIs." Read quickly, that sounds like model routing, which would make it a direct CometAPI competitor.

It is not, so far. The five catalog groups on its live home page are search, scraping, enrichment, agent actions and market and social data. No model-inference group appears in them. The Premium price band mentions "AI analysis", but the docs give no model-routing example. The YC text is also out of date elsewhere: it promises $10 in credits where the docs now say $5. We treat it as old copy, not a roadmap.

What would a third option look like?

The August version of this guide had no alternatives section for the agent-tool job. It needs one, because Orthogonal is not the only way to give an agent tools.

Composio is the clearest alternative we read on 7 November 2026. It advertises 1,500+ tools, a managed MCP gateway and a firewall feature that blocks instructions hidden in tool results. Its free Hobby plan includes 100,000 tool calls a month, and the Pro plan adds $29 in monthly credits.

Its model differs from Orthogonal's. Composio centres on connecting your users' own accounts (Gmail, Slack, Jira), where Orthogonal centres on paid data and search APIs billed per call. Composio does not have a Hokai record yet, so we do not link a page for it, and it is queued for one.

For a single job, you can also skip the gateway. If your agent only needs web search, a direct Tavily account is one fewer hop. For sending mail, see our email API guide for agents. For agent-ready servers, see our MCP servers roundup. And if you are comparing gateways as a class, the AI gateways directory lists every one we cover.

How does each one get set up?

Both start in a few minutes. The steps below come from each vendor's own quickstart and pricing text, so treat them as the happy path.

Who should pick which?

Match the tool to the job you have this month, not the label on the box.

• You route chat, code or media calls across vendors: CometAPI, for the flat 20% and one invoice.
• You build an agent that enriches leads, scrapes or searches: Orthogonal, for one key across 60+ providers.
• You need both reasoning and action: run both and keep two balances.
• You need your users to connect their own Gmail or Slack: look at Composio instead.
• You only use one or two APIs: skip both and buy direct.

What does a month of use cost?

Prices on a table tell you little until you put volume under them. The sums below use only the figures printed above, so they are illustrations, not quotes. Your mix of models and endpoints will change them.

The discount is a fixed share, so it scales with spend and never turns into a flat fee you have to cover first. Ten times the volume saves ten times as much. An outreach agent running all three Orthogonal lines spends nearly four hundred dollars on data before a single model token is paid for.

That last point is the one buyers miss. The two bills stack. The model meter pays for thinking, and the tool meter pays for fetching. If you cut the first by 20% through a router, the second does not move, so a data-heavy agent may see little change in its total. A chat product with no tool calls sees the full 20%.

Two habits keep the tool meter honest:

• run a small batch first and read the cost that every Orthogonal response returns, then multiply;
• check whether a provider such as Apollo.io or Tavily is cheaper bought direct once your volume is steady.

What does switching cost?

Moving off CometAPI means changing a base URL and swapping model name strings. Because the client is OpenAI-compatible, most integrations need nothing else. That low switching cost is also why its moat is thin: another router could match a 20% discount without much engineering.

Moving off Orthogonal costs more. Every call is wired to a specific request and response shape, so replacing it means rebuilding each integration against Apollo.io, Tavily or another provider directly, or finding another agent-tool gateway that wraps the same providers.

Two more frictions sit underneath. Credit balances on both services are pre-funded, so check refund terms before loading a large balance. And Orthogonal's catalog leans toward B2B data and search, so a team needing something outside that lane will not find it there yet.

What are the risks of putting a middleman in the path?

Every gateway adds a hop, and a hop adds failure modes. Be specific about which ones matter before you commit.

A model router sits between your app and the model maker. If the router has an outage, your product is down even when the maker is up. Your prompts also pass through a third party, which matters if they hold customer data. A discount this size comes from somewhere, and CometAPI's own page says it buys credits in bulk from several providers and routes requests across regions. That is a plausible model, but you are trusting the router's supply, not the maker's contract.

If you want a self-hosted alternative that keeps the hop inside your own network, LiteLLM is the usual starting point, and APIMODELS is a hosted second opinion.

A tool gateway has a different risk. It holds the accounts and the billing for every provider behind it, so a pricing change by one provider reaches your agent without a notice from that provider. Orthogonal returns a price with each search result, which softens this, but an agent that runs unattended can still burn through a $5 balance quickly if a loop calls a Premium endpoint. Set a spend cap in your agent code, not only in the dashboard.

Both services are pre-paid, so the failure that costs money is a stranded balance. Load small amounts until you have seen a full month of bills.

A coding agent such as Claude Code or Cursor makes the loop risk real, because it will happily retry a failing call. Put a hard ceiling on retries before you hand it a key.

When should you not buy either?

Skip CometAPI if you have a committed-spend contract with a model vendor that already beats 20% off, or if your compliance team requires a named subprocessor list and a signed data-processing agreement before any traffic flows. A 20% discount is not worth a procurement fight, and we found no published attestation to point your reviewers at.

Skip Orthogonal if your agent only ever needs two or three APIs. At that scale, direct accounts are cheaper to reason about than a new balance and a new vendor, and you keep the provider relationship. The gateway earns its place at around a dozen providers, where managing keys, billing and rate limits across accounts becomes the real cost.

Which one should you start with?

Most teams shipping a real agent product in 2026 will not choose. They will wire a model gateway for reasoning and a tool gateway for action into the same pipeline, because an agent that can only think is a chatbot, and an agent that can only run a fixed script is automation. Budget for both meters from day one.

If you want help picking the model side first, Smart Match takes a few questions about your use and returns a shortlist. If your agent runs on a frontier model such as Claude Fable 5, the gateway question is mostly about how you pay for it.

Sponsor spot open

What would change this verdict?

Watch two things over the next six months. First, whether Orthogonal adds a model-inference group to its catalog. Second, whether CometAPI ships anything resembling a tool layer of its own. The day either happens, "CometAPI vs Orthogonal" stops being a lesson about mislabelled categories and becomes a comparison worth running for real. We will re-read both pages then.

Frequently asked questions

What is the actual difference between CometAPI and Orthogonal?

CometAPI routes calls to 500+ foundation models, image and video generators through one OpenAI-compatible key. Orthogonal gives AI agents pay-per-call access to 60+ search, scraping and enrichment providers through an MCP server, an API and a CLI. They share the label gateway but almost no catalog.

How much does CometAPI cost compared to Orthogonal in 2026?

CometAPI prices official models at 0.8 times the provider rate, for example GPT 6.1 Terra at $0.80 per million input tokens against $1.00, with no monthly fee. Orthogonal charges per call from a pre-paid balance, from $0.001 to $1.00 depending on the endpoint, and gives new accounts $5 in free credit.

Can a team use CometAPI and Orthogonal together?

Yes, and most agent products do. CometAPI handles the reasoning layer by routing a call to a foundation model. Orthogonal handles the action layer by letting the same agent call search, scraping or enrichment APIs. Neither replaces the other, and each has its own balance.

Is Orthogonal really an AI model gateway like CometAPI?

Not on 7 November 2026. Its Y Combinator listing still names AI Model APIs as a starting area, but the live catalog groups are search, scraping, enrichment, agent actions and market and social data. No model-inference group appears among them.

Which one should a solo developer start with?

Start with CometAPI if the first thing you build is a chat or content feature that calls a model directly. Start with Orthogonal if you build an agent that must take an action or pull outside data, since its MCP connector adds to Claude or Cursor in one config step.

Covered in this guide

• CometAPI: CometAPI is a unified API gateway giving one key and one OpenAI-compatible endpoint to 500+ AI models at roughly 20% below official provider rates.
• Orthogonal: Unified API gateway for AI agents giving access to 30+ APIs via one key and MCP server, with $10 free credits and pay-per-call billing.
• APIMODELS: One API key for image, video, LLM and audio models, billed only on successful calls, with pay-as-you-go top-ups from $10 and no monthly fee.
• Apollo.io: Apollo.io is a sales intelligence platform with a searchable B2B contact database, built for prospecting and outreach with sequencing and a dialer attached.
• Claude Code: Anthropic's agent for the terminal, IDEs, desktop, web and mobile that edits files, runs tests and opens pull requests. There is no free plan.
• Claude Fable 5: Superseded by Fable 5.1 and Fable 5.2 but still sold on the Claude API: Anthropic's first generally available Mythos-class model, released June 2026, with a 1M-token context.
• Claude Opus 5: Anthropic's July 2026 Opus with a 1M token context and xhigh reasoning effort, still active but superseded by Claude Opus 5.5 on September 22, 2026.
• Cursor: Cursor is an AI code editor built on VS Code, widely adopted across large enterprise engineering teams, with Agent Mode, Tab completion, and Cloud Agents.
• DeepSeek-V4 Flash: Retired by DeepSeek on 10 September 2026; replaced by V4.1 Flash. Legacy API name now routes to the new model. MIT weights remain on Hugging Face.
• FLUX 3: FLUX 3 is a video, image, and audio model from Black Forest Labs, launched July 2026, built on a shared Self-Flow architecture.
• Gemini 3.6 Flash: Released July 21, 2026, Gemini 3.6 Flash is Google DeepMind's Flash-tier model built for agentic coding, computer use and long-context work.
• LiteLLM: Open-source AI gateway and Python SDK that puts 100+ LLM providers behind one OpenAI-compatible API, with budgets and fallbacks. Free to self-host.
• Tavily: Five-endpoint web access API for AI agents: search, extract, map, crawl and research, with a free plan and a keyless trial mode.

Sources