Managed inference for coding agents

Keep the reasoning. Control the run.

Prizmal manages inference for coding agents: curated open and frontier models, compute that adapts during generation, and policy enforced as the answer unfolds.

Keep your harness, prompts, and tools. We manage the models, the handoffs, and the record of each run.

Integration guide Explore the research

Agent task completion

84/100

Terminal-Bench v2.1

July 2026 · internally reported

Availability

99.9%

Monthly routing availability target.

Run controls

Policy during execution

Budgets, model permissions, and reviewable decisions.

Beyond intelligent routing

Serverless managed inference for coding agents. The control layer keeps deciding while the answer is written: open models carry the routine spans; frontier steps in only when the span requires it. We evaluate new models as they ship and update the mixture for your workload.

Illustrative overview. Not a measured trace.

Routers dispatch requests. Prizmal governs tokens: the decision happens while the answer is written, not before it.

DECISIONS WITHIN A RESPONSE

The Controller continues deciding during generation. Frontier capacity handles demanding spans, subject to the run policy.

COMPUTE INSIDE OPEN MODELS

RASP (Residual-Aware Sparsification) adapts compute inside open models for routine spans. Task constraints guide the managed run.

GOVERNED AT THE LAYER

Budgets and model permissions apply during execution. Operational metadata makes decisions reviewable.

Inside the Controller

Illustrative system view · model classes, not named models

Full topology. Requests, context and policy enter one boundary. The Controller assigns each span to a model class, the response returns to your harness, and every decision lands on the audit rail.

Anatomy of a managed run

Decisions continue within the response. Move the cursor or pick an event to inspect each span.

Illustrative system view · synthetic example · 512 tokens

Signal and allocation are an authored illustrative scenario, not allocation computed from the threshold.

Illustrative allocation, not a measured token stream.

Token position

The record contains operational metadata, not retained prompt or completion content.

Model a scenario

A quick look at what Prizmal managed inference could do for your workload

Employees (seats)

50 seats

Tokens per employee / day

66M / day

Illustrative workload assumptions. Adjust these inputs to match your usage.

Comparison blend / yr$577,500

Total with Prizmal / yr$199,125

Comparison baseline: $577,500
Prizmal total: $199,125
Platform ($50/seat/month): $30,000
Routing ($0.75 / 1k calls): $4,125
Modeled compute ($0.20/M): $165,000

Saving

$378,375

66% lower than the annual comparison.

Email this scenario

Assumptions: 250 working days/year, $50 per seat per month platform fee, $0.75 per 1,000 routed calls, 150,000 tokens per call. Modeled compute blend: $0.20/M tokens. Comparison blend: $0.70/M tokens. Blends are assumptions, not market quotes. Actual usage and mix are metered per agreement. The seat fee is not an all-in usage total. See Pricing.

Control the run as it happens

Budgets and model permissions apply during execution. Every decision is reviewable as operational metadata.

Illustrative system view · synthetic examplerun_sample_0417 · 512 tokens · 10 events

Permission gate

3 allowed · 1 restricted

Budget gate

Approaching limit from evt_07

Frontier tokens

23.4%

Sample value · 120 of 512

Records written

Metadata only, no prompt content

Model allocation strip

Open model 69.9% · Specialized open 6.6% · Frontier 23.4% · sample values

tok 0256512

Metadata ledger

Synthetic sample ledger, select a row for details
ts	tok	kind	model class	gate	outcome
T+0.00s	0	admit	Open model	budget	allowed
T+1.32s	110	escalate	Frontier	permission	allowed
T+1.80s	150	return	Open model	budget	allowed
T+2.35s	196	route	Specialized open	permission	allowed
T+2.76s	230	return	Open model	budget	allowed
T+3.22s	268	escalate	Frontier	permission	allowed
T+3.72s	310	return	Open model	budget	allowed
T+4.82s	402	escalate	Frontier	permission	restricted
T+5.28s	440	return	Open model	budget	allowed
T+6.14s	512	complete	Open model	record	recorded

Sample export · 10 records

{
  "note": "Illustrative system view · synthetic example",
  "records": [
    {
      "synthetic": true,
      "run_id": "run_sample_0417",
      "event_id": "evt_01",
      "ts": "T+0.00s",
      "token": 0,
      "kind": "admit",
      "model_class": "open",
      "rasp": true,
      "gate": "budget",
      "outcome": "allowed",
      "budget_state": "within"
    },
    {
      "synthetic": true,
      "run_id": "run_sample_0417",
      "event_id": "evt_02",
      "ts": "T+1.32s",
      "token": 110,
      "kind": "escalate",
      "model_class": "frontier",
      "rasp": false,
      "gate": "permission",
      "outcome": "allowed",
      "budget_state": "within"
    },
    {
      "synthetic": true,
      "run_id": "run_sample_0417",
      "event_id": "evt_03",
      "ts": "T+1.80s",
      "token": 150,
      "kind": "return",
      "model_class": "open",
      "rasp": true,
      "gate": "budget",
      "outcome": "allowed",
      "budget_state": "within"
    },
    {
      "synthetic": true,
      "run_id": "run_sample_0417",
      "event_id": "evt_04",
      "ts": "T+2.35s",
      "token": 196,
      "kind": "route",
      "model_class": "specialized",
      "rasp": false,
      "gate": "permission",
      "outcome": "allowed",
      "budget_state": "within"
    },
    {
      "synthetic": true,
      "run_id": "run_sample_0417",
      "event_id": "evt_05",
      "ts": "T+2.76s",
      "token": 230,
      "kind": "return",
      "model_class": "open",
      "rasp": true,
      "gate": "budget",
      "outcome": "allowed",
      "budget_state": "within"
    },
    {
      "synthetic": true,
      "run_id": "run_sample_0417",
      "event_id": "evt_06",
      "ts": "T+3.22s",
      "token": 268,
      "kind": "escalate",
      "model_class": "frontier",
      "rasp": false,
      "gate": "permission",
      "outcome": "allowed",
      "budget_state": "within"
    },
    {
      "synthetic": true,
      "run_id": "run_sample_0417",
      "event_id": "evt_07",
      "ts": "T+3.72s",
      "token": 310,
      "kind": "return",
      "model_class": "open",
      "rasp": true,
      "gate": "budget",
      "outcome": "allowed",
      "budget_state": "approaching"
    },
    {
      "synthetic": true,
      "run_id": "run_sample_0417",
      "event_id": "evt_08",
      "ts": "T+4.82s",
      "token": 402,
      "kind": "escalate",
      "model_class": "frontier",
      "rasp": false,
      "gate": "permission",
      "outcome": "restricted",
      "budget_state": "approaching"
    },
    {
      "synthetic": true,
      "run_id": "run_sample_0417",
      "event_id": "evt_09",
      "ts": "T+5.28s",
      "token": 440,
      "kind": "return",
      "model_class": "open",
      "rasp": true,
      "gate": "budget",
      "outcome": "allowed",
      "budget_state": "approaching"
    },
    {
      "synthetic": true,
      "run_id": "run_sample_0417",
      "event_id": "evt_10",
      "ts": "T+6.14s",
      "token": 512,
      "kind": "complete",
      "model_class": "open",
      "rasp": true,
      "gate": "record",
      "outcome": "recorded",
      "budget_state": "approaching"
    }
  ]
}

Audit records contain operational metadata, not retained prompt or completion content. Operational audit metadata has separate retention.

Read about data handling and controls →

We run the AI stack. You ship your product.

Consistent reasoning means the run keeps its task constraints, holds up under comparable evaluation, and stays inside your policy from first token to last.

Prizmal manages model selection, compute, and handoffs. Your team keeps its agent harness, prompts, and tools.

Illustrative system view · synthetic exampleevaluation workspace · generic candidates

Mixture admission. Admit the candidate to the workload mixture for the span types it handled well.

Evaluation matrix. Select a candidate for details.
Task completion	Constraints	Tool behavior	Serving config	Policy review	Stage
pass	pass	pass	pass	pass	Observe runs
pass	pass	review	pass	pass	Mixture admission
pass	pass	pass	pass	review	Policy review
review	fail	pending	pending	pending	Evaluate

● pass · ◐ review · × fail · ○ pending · qualitative gates, no scores

How it works

Three steps from your current stack to managed inference

01
Request access
Tell us about your setup. Onboarding is manual; Prizmal provisions provider access with one Prizmal key.
02
Configure the supported client
Use the client example below with your Prizmal key and model prizmal-auto. Check current client requirements.
03
Inspect a first run
Send a small test request. Inspect the response, tool behavior, and trace before using it in your workflow.

1# ~/.zshrc or ~/.bashrc
2export ANTHROPIC_BASE_URL=https://api.prizmal.ai/anthropic
3export ANTHROPIC_AUTH_TOKEN=$PRIZMAL_API_KEY
4export ANTHROPIC_MODEL=prizmal-auto

Claude Code appends /v1/messages to the base URL. Anthropic does not support routing Claude Code to non-Claude models; confirm Prizmal model and routing scope before deployment.

One Prizmal key; Prizmal provisions provider access. Check your client's docs for current config keys and supported features. PRIZMAL_API_KEY is a placeholder for your provisioned key.

Client configuration and supported features vary. Harness support matrix

Client setup walkthrough

One key, one base URL, first routed request. See the full developer setup →

Keep your agent setup. Let us manage inference.

One Prizmal key. Curated models, adaptive compute, and policy during execution. We onboard teams by hand.

Integration guide

Prompts and completions are not retained after serving. Operational metadata has separate retention. How we handle data