Reference
Limits, caps and rate limits
Every hard boundary in one place — sequence caps, upload caps, per-endpoint rate limits, retention windows and tier differences. Values are read from the running engine, not restated from memory.
Updated 31 July 2026
All documentation
On this page
Knowing where a tool stops is the difference between using it and misusing it.
Sequence and file caps
| Limit | Value | Applies to |
|---|---|---|
| CRISPR input sequence | 1 Mbp | Paste or resolved gene. Largest human genes reach ~2.4 Mbp; at that scale, paste the region you are editing |
| CRISPR minimum | 23 nt (SpCas9), 27 nt (Cas12a) | Below this no guide can fit with its PAM |
| Upload or paste | 8 MiB | Directed evolution and plasmid import |
| Request body | 25 MB | Server-wide, covers the largest legitimate payload |
| Primer template | 60,000 bp | Primer analysis and in-silico PCR |
| Primers per run | 50 | Analysis and multiplex |
| Primer length | 10–60 nt | Design bounds |
| PCR product | 4,000 bp | Products above this will not amplify under standard conditions |
| Design flank | 400 bp | How far either side of a target the designer hunts for a primer |
| Local specificity genome | 12 MB file, 30 Mbp scanned | Bacterial and small eukaryotic genomes |
| Round 2 / epistasis protein | 500 aa | Bounds ESM scoring time on shared CPU |
| Agent message | 4,000 characters | Per message to Turing |
| ESM-2 context | 1,022 residues | Longer proteins are scored in overlapping windows |
Rate limits
Per IP, sliding window, counted independently per endpoint prefix. Exceeding one
returns 429 with a Retry-After header naming the seconds to wait.
| Endpoint prefix | Requests | Window |
|---|---|---|
/api/orchestrator/poll |
240 | 60 s |
/api/collect, /api (default) |
120 | 60 s |
/api/plasmid, /api/ping, /api/orchestrator/runs |
60 | 60 s |
/api/preview, /api/primers/analyze, /api/crispr |
40 | 60 s |
/api/primers/design, /api/primers/multiplex, /api/structure, /api/project |
30 | 60 s |
/api/run, /api/identify, /api/primers/local-specificity |
20 | 60 s |
/api/orchestrator (start, reply, stop) |
12 | 60 s |
/api/de/epistasis |
8 | 60 s |
The tight budgets are the expensive ones: /api/run is a full ESM-2 pipeline
and /api/de/epistasis can be a dozen forward passes.
You will not hit these by hand
Every limit above is far above manual use. They exist to bound scripted traffic on shared CPU, not to ration the interface.
Accounts and tiers
| Anonymous | Free account | Pro | |
|---|---|---|---|
| Chat with Turing | Yes | Yes | Yes |
| Run a tool via Turing | 5-minute trial | Yes | Yes |
| Directed evolution | 5-minute trial | Yes | Yes |
| CRISPR, primers, plasmid | No — sign-in required | Yes | Yes |
| Save designs and libraries | No | Yes | Yes |
| Artifact retention | — | 45 days | No expiry |
| ESM-2 model | 35M | 35M | 35M today; 650M and 3B are labelled “Pro · GPU (coming soon)” and are not yet wired |
The anonymous trial is time-based, not run-based: five minutes from your
first run, tracked in a signed cookie. It covers /api/run and the agent only.
CRISPR, primer analysis and the plasmid editor have no trial and return 403
with a sign-in prompt.
The 650M and 3B model options are visible but disabled
They appear in the model picker marked “Pro · GPU (coming soon)”. They are not available on any tier today, and no benchmark figures exist for them.
Retention and expiry
| Thing | Lifetime |
|---|---|
| Upload session | 6 hours |
| Agent run | 6 hours |
| Free-tier saved library | 45 days |
| Pro saved library | No expiry |
In-flight work does not survive a deploy
The engine runs a single worker process, so job state, upload sessions, the rate-limit counters and the loaded model all live in memory. A deploy or restart ends every running job and agent run. Polling an expired run returns 410; the chat silently starts a fresh one. Saved designs are in the database and are unaffected.
Coverage limits
| Limit | Detail |
|---|---|
| Genome off-target search | Only the top 10 ranked guides are screened — each genome query costs a few seconds |
| Human and mouse indexes | Coding sequence only. Intronic, intergenic and regulatory off-targets are outside the index |
| E. coli, yeast, C. elegans, D. melanogaster | Complete genome, no coding-only caveat |
| Index build | Roughly one second per megabase, on first use per organism |
What the engine does not do
- On-target activity is a heuristic in the style of Doench et al., not a re-trained or independently benchmarked model.
- Base-editing windows are literature consensus, not a trained efficiency predictor.
- Cloning, digests and in-silico PCR are predictions. Confirm overhangs, junctions and orientation before committing reagents.
- Directed evolution is not de novo design. It improves an existing protein along an existing axis.
- Primer analysis ranks primers you supply. The separate design tool proposes candidates from a template; the analysis tool does not invent them.
- There is no programmatic API access. Authentication is a browser session token; no API keys are issued.
Research use only
Not for clinical, diagnostic, therapeutic, prophylactic, food, feed or cosmetic use. See Terms §12 and the Biosecurity Policy. Every export carries this statement.