OpenAI Amp Agents Beyond Code — Pricing Impact (Oct 2026)
Amp uses OpenAI models for planning, writing, design, and voice. See ChatGPT billing exclusions, Plaid's 6× token cost, and current API prices.
By AI Pricing Guru Editorial Team
AI Pricing Guru articles are maintained by the editorial workflow behind the site: daily pricing snapshots, provider source checks, and review passes for model launches, subscription limits, and billing changes.
TL;DR
- OpenAI's October 14 Amp story covers agents doing planning, documentation, design, marketing, and voice-led coordination. It announces no model release or price cut.
- Linking ChatGPT moves eligible OpenAI model usage off Amp credits, not every charge. Orbs, voice, search, webpage reading, and image generation remain separate.
- Amp's Plaid uses OpenAI Ultrafast: up to 6× faster inference at 6× token cost. It requires Amp-provided inference, not a linked ChatGPT subscription.
- Existing ChatGPT subscribers should check routing and plan limits first; buy premium speed only when measured time savings justify it.
Amp pricing: subscriptions, cloud runners and premium inference
Official sources checked ; daily verification watch. Amp plan source.
| Tier | Monthly price | Orb allowance and scope |
|---|---|---|
| Hobby | $0.00 | Own runners or pay-as-you-go orbs; upstream tokens and tools separate |
| Megawatt | $20.00 | 45,000 published minutes. Published orbin time allowance; size-weighted consumption, not identical wall time for every orb |
| Gigawatt | $200.00 | 480,000 published minutes. Published orbin time allowance; size-weighted consumption |
| Teams | No extra workspace charge | No extra workspace charge; members choose individual tiers; pooled credits |
| Enterprise | Custom quote | Custom quote; pooled credits, no paid user seats |
Subscription/BYOK provider limits still apply. Voice, search, webpage reading and image generation remain separate Amp-credit charges. Orb allowances are size-weighted; included usage resets monthly without rollover.
| Processing tier / context | Input / cached / cache write | Output |
|---|---|---|
| Standard · short context | $10.00 / $1.00 / $12.50 | $50.00 |
| Ultrafast / Plaid · short context | $60.00 / $6.00 / $75.00 | $300.00 |
| Standard · long context | $20.00 / $2.00 / $25.00 | $75.00 |
| Ultrafast / Plaid · long context | $120.00 / $12.00 / $150.00 | $450.00 |
Above 272,000 input tokens, long-context rates apply to the full request. Official rate card. Amp-provided inference only; not linked ChatGPT; subagents and unsupported models fall back to Fast or Standard. Up to 6× inference speed is a vendor claim, not a measured end-to-end task result.
| Size | CPUs / RAM | Standard rate / hour |
|---|---|---|
| a1.tiny | 1 / 2 GB | $0.08 |
| a1.small | 2 / 4 GB | $0.17 |
| a1.medium | 4 / 8 GB | $0.33 |
| a1.large | 8 / 16 GB | $0.66 |
| a1.xxlarge | 16 / 32 GB | $1.32 |
| a1.3xlarge | 16 / 44 GB | $2.13 |
Billed by minute; paused orbs cost nothing. The largest size requires Gigawatt or Enterprise, whose rates are project-specific. Orb billing rules · Maintained structured source record.
GPT-Live-1 direct voice rate—not an all-in Puck bill
| Model | Primary / min | Primary / hour | Streaming / hour | Change |
|---|---|---|---|---|
| GPT-Live 1 | $0.05 | $3.00 | — | Baseline |
Primary rates use the provider's published minute and hour values. Speech-to-text rows show REST as primary and list the separate WebSocket streaming rate where available. Additional text, tools, or telephony charges may apply.
Live values come from the canonical pricing API.
Direct API baseline—not Amp credits or ChatGPT allowances
USD per 1M tokens. Input and output rates are charted separately.
Estimate a measured text-agent workload
Assumes 75% input tokens and 25% output tokens using current per-million rates.
GPT-6.1 Sol
openai
$40.00
- Input share
- $15.00
- Output share
- $25.00
GPT-5.6 Terra
openai
$45.00
- Input share
- $15.00
- Output share
- $30.00
GPT-6 Astra
openai
$200.00
- Input share
- $75.00
- Output share
- $125.00
Current Standard API rates for Amp's ChatGPT Only models
| Model | Provider | Input / 1M | Cached / 1M | Output / 1M |
|---|---|---|---|---|
| GPT-6 Astra | openai | $10.00 | $1.00 | $50.00 |
| GPT-6.1 Sol | openai | $2.00 | $0.1 | $10.00 |
| GPT-5.6 Terra | openai | $2.00 | $0.2 | $12.00 |
Built from pricing.json at publish time.
Amp is taking agents beyond code; the cost question is which work your subscription actually covers. OpenAI’s October 14 customer story describes broader workflows, not a new unlimited-agent plan.
What changed: planning, design, and voice-directed agents
Amp CEO Quinn Slack says OpenAI models account for the majority of Amp’s tokens. The story describes agents deciding what to build, producing documentation and pages, and continuing work after launch. Amp’s 20-person team also uses an agent for marketing and new-user experience rather than assigning someone exclusively to marketing.
GPT-6 Astra supports writing and design; Slack’s personal example uses house plans to create a Blender model. Puck coordinates agents and retrieves earlier conversations. Calling Puck uses GPT-Live-1 for real-time spoken interaction, including from a watch.
These are customer-reported examples, not a controlled productivity benchmark. The story supplies no acceptance rates, invoice, or quantified savings. Slack describes some users consuming 10 billion tokens a week; it does not identify their model mix, cache hits, billing routes, or output share. That number cannot establish a representative weekly bill.
Pricing impact: three different billing paths
Amp’s current Dial documentation provides the actionable detail behind “Sign in with ChatGPT.”
Linked ChatGPT: Eligible OpenAI inference is billed to the linked subscription instead of Amp credits. Linking alone does not convert every role: non-covered models can still consume credits. The ChatGPT Only preset routes low, medium, and high roles to OpenAI models: Terra for the low main agent, Sol for medium, and Astra for high and every Oracle. Subagent routing also follows the preset. Ultra retains Amp’s own routing.
Separate Amp charges: Orbs—Amp’s cloud development environments—are not included in ChatGPT. Voice, web search, webpage reading, and Painter image generation use Amp credits. Using a runner on your own machine avoids Amp orb charges, but not your machine’s operating costs. Consult Amp’s pricing page for current tiers. “No Amp token fees or limits” is not evidence of unlimited upstream subscription usage or free tools.
Premium speed: Amp’s September 29 Plaid announcement says Astra inference can run up to 6× faster at 6× cost per token. Plaid works only with Amp-provided inference; linked ChatGPT subscriptions do not support it. Subagents and unsupported models fall back to Fast or Standard. Faster inference is not necessarily sixfold faster task completion when tools and review dominate.
The generated table compares the preset’s models at direct OpenAI Standard, short-context API rates. It is not an Amp invoice or a subscription allowance. The OpenAI pricing page covers cache writes, service tiers, and higher rates above 272,000 input tokens. Use the token calculator with actual usage; its text estimate excludes voice, images, orbs, and other tools.
Labs availability and explicit workflow blockers
Our Labs coverage notes record the access boundary: Astra is unavailable on the maintained benchmark route, and Amp/Puck/Plaid has no authorized, pinned replay in our suite. The existing exact GPT-6.1 Sol text result is not a score for writing, visual design, marketing or full-duplex voice. A valid comparison needs fixed tasks and tools, raw outputs, timing, complete subscription/credit/token/orb records, and independent accepted-output grading. No customer claim becomes a Labs score.
What this means—and what to do now
Existing ChatGPT subscribers benefit if eligible agent work moves off Amp credits without exhausting their plan. Teams doing writing and design benefit from broader workflows, but still need to measure editing time and accepted outputs. Budget-sensitive users can lose the expected saving when premium speed, tools, or cloud environments dominate.
- Inspect every role. Apply ChatGPT Only if appropriate, then check main-agent, Oracle, and subagent routing. Confirm your plan’s eligibility and remaining usage.
- Pilot one non-code task. Test a documentation page or product brief. Record tokens, Amp credits, retries, and reviewer time per accepted result.
- Separate the meters. Track subscription usage, orb time, voice, search, images, and Amp-funded inference independently.
- Test speed economically. Compare Standard with Plaid on the same task; pay for it when reduced waiting outweighs the token premium.
Compare Anthropic pricing for another provider baseline and read our API cost guide. For a separately built, bounded drafting workflow, test a hosted open-model baseline on Novita—not a replacement for Amp’s integrated workflow or proof of equal quality.
Affiliate disclosure: This article contains an affiliate link. We may earn a commission at no extra cost to you.
Sources: OpenAI’s customer story, official API pricing, GPT-Live-1 docs, Amp pricing, orb billing, Dial and Plaid documentation, checked October 14, 2026. Prices render from maintained canonical records; daily pricing and official-source change monitoring cover this page. Voice sessions are billed per second; backend model/tool usage is separate, so the direct voice table is not a retail Puck invoice.