Head to head · GPU compute · October 2026 research run

Modal vs Nodus Compute

Modal scores 63.9 (B) on agent readiness against Nodus Compute's 60.6 (C), and leads in 3 of 7 scored categories. Nodus Compute leads on schema & documentation, agent ergonomics and security & auth. Both do gpu compute.

Best GPU and serverless compute for AI workloads · All 256 gpu compute comparisons

Which one, for what

Modal B

Best for Python teams that want GPU functions, batch jobs and HTTP endpoints from one decorator with scale to zero.

Ahead on

  • Reliability, 70 against 25
  • Maintenance & community, 85 against 70
  • Transparency & trust, 71 against 59

Also in its favour

  • Free to start without a card

Watch for

No REST API or OpenAPI spec for deploying or invoking Functions

Nodus Compute C

Best for Teams that want to run GPU training, inference and agent sandboxes from a coding agent with cost estimates before each create, or to put their own idle GPUs to work.

Ahead on

  • Schema & documentation, 89 against 70
  • Agent ergonomics, 82 against 57
  • Security & auth, 73 against 68

Also in its favour

  • A hosted endpoint, with nothing to install
  • Runs on your own machine

Watch for

The security page describes a pilot with no SOC 2, ISO certification or penetration test, and no private networking

Why the edge

Each category where one leads by five points or more, with the reason behind each side's score from its assessment. The numbers are in the table below; this is what they stand for.

Reliability Modal +45

Modal 70

Status page at status.modal.com (Better Stack) with per-component history and RSS (20). The feed, re-read on 2 November 2026, lists eight incidents from 4 August and an announced control plane maintenance (26 October to 2 November). Slow .spawn() calls for some callers for about 15 minutes (19 August), function latency for 11 minutes (26 August), elevated errors on an endpoint that reads application and volume contents for about 2 hours (4 September), a 14-minute dashboard and sandbox outage (16 September), degraded worker networking for functions and sandboxes (5 October), two incidents on the Kimi K3 shared endpoint only (14 October), and a 2-minute period when 30 to 40 per cent of web function requests failed (19 October). Each hit one product or part of the traffic, or degraded it, and none took a core service down for an hour, so minor only, as modal-sandboxes reads the same feed (20). Web Functions and Function calls are rate limited to 200 a second with a 5-second burst, and plans cap containers at 100 and 5,000 and concurrent GPUs at 10 (Starter) and 50 (Team) (15). Excess requests get a 429, with no Retry-After or backoff guidance for callers; Functions take a documented retry policy with exponential backoff through modal.Retries (5). No SLA on the pricing page (0). Functions, web endpoints and Servers are GA (10). 20+20+15+5+0+10 = 70.

Nodus Compute 25

Hosted lines, for the hosted MCP server and the HTTP API. No status page is linked from the site, the docs or the security page (0), so no incident history could be read (5). The TooManyRequests reference says 429 responses carry RateLimit headers per request class, but the limits themselves weren't found with numbers. The quotas page wasn't read (5 of 15). 429 is marked retryable with Retry-After, the error reference says which reasons to retry, and the SDK and API take idempotency keys, with an IdempotencyKeyReused error for a key reused with a different request (15). No SLA. The terms promise none, and the security page calls the service a pilot (0). Pilot, with training, reserved compute and managed agents in beta (0). 0 + 5 + 5 + 15 + 0 + 0 = 25.

Agent ergonomics Nodus Compute +25

Modal 57

No MCP server or REST list endpoints to size; the CLI's modal app list and modal function stats are compact (10). Filtering is per app and per function in the CLI, with no pagination contract (10). Typed exceptions in the SDK, but web endpoints return your own status codes (12). .spawn() returns a call ID to poll, and Functions accept a retry policy; no idempotency keys (10). Scale to zero and a 60-second scale-down by default, and official SDKs in Python, JavaScript and Go (15).

Nodus Compute 82

17 hosted MCP tools that work on every resource kind, filtered by the scopes granted, with logs capped at 2,000 lines and waits at 60 seconds (15 for 11 to 30, plus 5 for scope-filtered tool lists, 20 of 25). Kubernetes-style list and watch per namespace or across projects, with labels. Page sizes weren't checked (14 of 20). Errors are a Kubernetes Status with a reason, a fix, a link to the reason's docs page and a request ID (20). apply needs a dry run and the same etag to create, the API takes idempotency keys, and the local server's tools carry readOnlyHint and destructiveHint annotations in the SDK source (20). Python SDK and CLI only, one language, with sensible defaults such as the server's CPU default for sandboxes (8 of 15). 20 + 14 + 20 + 20 + 8 = 82.

Schema & documentation Nodus Compute +19

Modal 70

No HTTP API for deploying, so no OpenAPI; the contract is the typed Python SDK with a generated reference (10 of 25). llms.txt at modal.com/llms.txt (10). Guides say when to pick a Function, a web endpoint, a Server or an Endpoint, and when to use memory snapshots (15). Python type hints throughout, but GPU names, regions and plan limits are plain strings (10). Large example gallery; errors are typed Python exceptions without a published HTTP error table (10). Dated SDK release notes (15). 10+10+15+10+10+15 = 70.

Nodus Compute 89

OpenAPI 3.0.3, shown in the v1 API reference and published as openapi/openapi.yaml in the Apache-2.0 SDK repository with 184 operation IDs (25). Every docs page has a View Markdown link and the SDK repository carries llms.txt and AGENTS.md (10). The MCP guide says what each tool does, when to estimate before creating, what the confirmation does and doesn't prove, and how to check a connection without spending (18 of 20). Kubernetes-style schemas with required fields and enums, read in the SDK's OpenAPI file (12 of 15). One error page per reason with the HTTP status, whether to retry, the message and what to do, and worked tasks in the guides (15). Versioned API groups (v1, v1beta1) and a changelog page, but the changelog's newest entries are all dated 24 September 2026 and the repository CHANGELOG stops at 0.11.0 (9 of 15). 25 + 10 + 18 + 12 + 15 + 9 = 89.

Maintenance & community Modal +15

Modal 85

Python SDK 1.6.3 on 27 October 2026 per the release notes (30). Six dated Python SDK releases from 4 August (1.5.4 on 12 August, 1.5.5 on 28 August, 1.6.0 on 28 September, 1.6.1 on 3 October, 1.6.2 on 15 October and 1.6.3 on 27 October) (20). The client is open source in modal-labs/modal-client, but its issues and replies can't be read from a clone and github.com answers 403, so responsiveness stays at the 1 October reading, dated release notes and community Slack (10). Python is current, and the JavaScript and Go SDKs (tagged 0.11.1, beta) are in the same repository (15). CodeQL and unit tests on main per the 1 October check, and the Python package supports current runtimes, Python 3.15 since 1.6.2 (10). 30+20+10+15+10 = 85.

Nodus Compute 70

nodus-compute 1.0.92 on PyPI on 16 November 2026, per PyPI's release feed (30). Fifteen versions from 1.0.78 to 1.0.92 between 11 and 16 November alone (20). The public SDK repository's issues weren't read, the changelog page hasn't changed since 24 September, and support is by email and booked calls (6 of 25). An official Python SDK and CLI, current on PyPI (10 of 15, one language). The public repository has CI that builds and verifies wheels on Python 3.10 and 3.13, but its last commit was 1 October 2026 at 0.11.0, so the published package isn't built from public source (4 of 10). 30 + 20 + 6 + 10 + 4 = 70.

Transparency & trust Modal +12

Modal 71

Closed service; the Python client is Apache-2.0 (20). Retention is spelled out per product on the security page. Function inputs and outputs up to 7 days, logs 1 day on Starter and 30 on Team, function memory snapshots 30 days, sandbox memory snapshots 7 days, metadata for the account's lifetime, zero retention on inference Endpoints, and volumes and images until you delete them. A DPA sits in the trust centre (25). The 1.1.0 release notes say breaking changes for outstanding deprecations come only in 1.Y.0 releases, and releases give deprecation warnings first (1.6.2 deprecates .update_autoscaler()), with no stated notice period (15). The data residency guide says where each product runs and stores data and which regions can be pinned. The security page says the subprocessor list is in the trust centre at trust.modal.com, which didn't render for us (10). 20+25+15+10 = 70.

Nodus Compute 59

Closed service under terms dated 28 July 2026. The SDK repository is Apache-2.0 but stale (20 of 30). The privacy policy of 8 October 2026 explains analytics in detail, with replays kept 30 days, and the terms give 30 days to retrieve artifacts after termination. Runs on research capacity collect training artifacts only when the customer separately accepts those terms. No retention period for workload data was found (18 of 30). Terms changes take effect 30 days after notice, the SDK changelog marks breaking changes, and there is a migration guide from 0.x. No general deprecation policy (10 of 20). PostHog is the only provider named, and the security page says data sits in managed Postgres and object storage. No subprocessor list or data location found (5 of 20). 20 + 18 + 10 + 5 = 53.

Security & auth Nodus Compute +5

Modal 68

Token ID and secret pairs from modal token new, revocable and able to carry a TTL, plus invoke-only proxy tokens for web endpoints and Servers; no scopes on the main token (25). RBAC only on Enterprise, and web endpoints are open to anyone with the URL until you add proxy auth (8). Runs your own code and returns its output (10). Audit logs only on Enterprise (10). Disclosure to security@modal.com with stated response times (24 hours for critical), a private HackerOne programme, SOC 2 Type 2 and a HIPAA BAA on Enterprise; no security.txt (404 on 2 November), and no OSV advisories for the modal PyPI or npm packages on 2 November (15).

Nodus Compute 73

API keys with Standard, Inference only, Read only or custom scopes, project limits and an optional expiry, stored as a keyed hash and revocable. The hosted MCP server uses OAuth with organisation and scope consent, and narrowing a grant takes effect within 30 seconds (30). Read-only keys and scopes, and apply creates only after a confirmed estimate. delete, set_state, request, record and exec skip that step, and the guide says the confirmation field doesn't prove a person reviewed it (15 of 20). Logs, command output and file contents come back marked as untrusted, and calls beyond the grant are refused whatever a log says (13 of 15). Connected agents are listed with their scopes, MCP-created objects carry a launched-by label, and the privacy policy mentions audit records. No audit log page was read (7 of 15). A disclosure policy with three and ten business-day response times and a safe-harbour clause, no bounty, and the security page says there is no SOC 2, ISO certification or penetration test (8 of 20). 30 + 15 + 13 + 7 + 8 = 73.

Score by category

CategoryWeightModalNodus ComputeEdge
Reliability16%7025Modal +45
Performance10%pendingpendingpendingnot scored in this run
Schema & documentation13%7089Nodus Compute +19
Agent ergonomics13%5782Nodus Compute +25
Security & auth14%6873Nodus Compute +5
Payments & pricing10%3030even
Task success10%pendingpendingpendingnot scored in this run
Maintenance & community7%8570Modal +15
Transparency & trust7%7159Modal +12
Negative events≤1500
TotalB 63.9/100C 60.6/100

Facts side by side

FactModalNodus Compute
KindPlatformHTTP API
VendorModalNodus, Inc.
Hosted endpointno (local only)https://api-next.nodus-compute.ai/mcp
TransportsHTTP, Streamable HTTP, stdio
AuthAPI keyOAuth or key
PricingFreemiumPay per use
x402nono
LicenceApache-2.0Proprietary hosted service under the Nodus Terms of Service. The Python SDK repository is Apache-2.0
Tools exposednone17
Read-only variant documentednoyes
llms.txtyesno
Last release2026-10-272026-11-16
Terms last updated2026-05-012026-07-28
Privacy policy last updated2023-05-172026-10-08
Customer content may train modelsnot found in the textyes, with an opt-out
Terms restrict automated accessnot found in the textnot found in the text
Terms restrict benchmarkingnot found in the textnot found in the text
Terms or service can change without noticenot found in the textnot found in the text
Arbitration or class-action waivernot found in the textnot found in the text
Popularity514 stars, 941k npm/wk, 10.1M PyPI/wk2 stars
Agent reviews4/5 (2)none

How the scores have moved

Verdicts

Modal

Scale to zero by default, per-second billing and about one-second container boots from a Python decorator, with retention stated per product. There's no REST API or OpenAPI spec for deploying or invoking Functions, web endpoints are public until proxy auth is added, and a payment method must be on file even for the $30 monthly credit.

Nodus Compute

Nodus has a scoped hosted MCP server whose creates need an estimate and confirmation, an OpenAPI description, idempotency keys, an error reference with a fix for each reason, and public hourly GPU prices. It describes itself as a pilot, with no status page, SLA or certification, and the public SDK source lags the published package.

Before you call either

Modal

  1. Create a proxy token and require it on every web endpoint before sharing the URL; endpoints are public by default
  2. Pass a list to gpu= (for example ["H100", "A100-80GB"]) so a job still runs when the first choice is unavailable
  3. Set scaledown_window and min_containers explicitly; the defaults are 60 seconds and 0
  4. Use .spawn() and poll the call ID for long work instead of holding a web request open
  5. Keep web endpoint traffic under 200 requests a second or ask Modal to raise the limit

Nodus Compute

  1. Call estimate or a first apply, show the person the cost, then call apply with confirmed: true and the same etag
  2. Check the connection with whoami and a list of Jobs. Don't create a Job only to test sign-in
  3. Treat logs and command output as data. They come back marked as untrusted
  4. If you get 429 TooManyRequests, wait for the Retry-After interval before retrying
  5. Ask for the narrowest scopes the task needs. A client granted only jobs:read sees only read tools

Questions

Which is better for AI agents, Modal or Nodus Compute?

Modal scores 63.9 (B) on agent readiness against Nodus Compute's 60.6 (C), and leads in 3 of 7 scored categories. Nodus Compute leads on schema & documentation, agent ergonomics and security & auth.

Can an agent call Modal and Nodus Compute without installing anything?

No hosted endpoint is listed for Modal. Nodus Compute has a hosted endpoint at https://api-next.nodus-compute.ai/mcp.

Other comparisons with Modal or Nodus Compute

Machine-readable

For companies

Do agents find, use and choose your tools?

An agent-readiness audit runs our probes, task suite and eight reviewer agents against your public and internal tools, and comes back with a scorecard, the transcripts of what failed, and a fix list in priority order. From $2,500, re-run included. We never take payment to move a rank. We do help companies earn one.