OpenAI GPT-5.6 Luna & Terra Price Cuts on Amazon Bedrock (July 2026)

AWS lifecycle notice (June 30, 2026) — Amazon Bedrock Agents Classic is in maintenance for new customers after July 30, 2026. Net-new agent builds should use Bedrock AgentCore. Full matrix: lifecycle roundup.

Update (September 22, 2026): GPT-6 Sol and GPT-6 Luna are GA on Bedrock (up to 1M context, explicit prompt caching, lower API pricing than these GPT-5.6 pins). GPT-6 Astra (September 8, 2026) is the flagship above them. This post stays the July 30 rate-card record. Routing: GPT-6 Sol and Luna on Bedrock. GPT-6.1 Sol followed on September 29, 2026 (US Geo CRIS only, $2.20 / $11.00 per 1M tokens on the model card).

On July 30, 2026, AWS announced that on-demand inference prices on Amazon Bedrock for OpenAI GPT-5.6 Luna dropped 80% and GPT-5.6 Terra dropped 20%, in line with OpenAI’s first-party price cut. GPT-5.6 Sol is unchanged. The new rates apply automatically.

This is a routing and FinOps event, not a “flip the default model ID” event. If you already run OpenAI on Bedrock (see our April 2026 OpenAI-on-Bedrock analysis), recalculate cost per completed task by lane — then promote only the lanes that still pass quality.

What changed on the Bedrock rate card

Published Amazon Bedrock on-demand rates after the cut (US East N. Virginia / Ohio and US West Oregon where listed):

Source: Amazon Bedrock pricing (OpenAI frontier table, post–July 30, 2026). OpenAI’s global API list is slightly lower ($0.20/$1.20 Luna; $2/$12 Terra); Bedrock’s published in-region figures track OpenAI’s data-residency tier (~10% uplift). Treat the Bedrock page as the budget source of truth for AWS bills.

Implied prior Bedrock-equivalent rates (from the stated 80% / 20% cuts): Luna ≈ $1.10 / $6.60, Terra ≈ $2.75 / $16.50 per 1M input/output.

AWS positions the tiers the same way it did at GPT-5.6 GA:

• Luna — fast, high-volume: classification, summarization, routing, customer-service automation, routine tool workflows
• Terra — everyday production with stronger reasoning at GPT-5.5-class quality, lower cost
• Sol — flagship multi-step / agentic reasoning (price unchanged — do not expect this cut to move Sol TCO)

On August 3, 2026, AWS expanded Sol, Terra, and Luna to 1 million token context windows on Bedrock. Prompt caching with explicit cache breakpoints applies to those long-context requests, so repeated prefix is billed at the cache-read discount rather than full input. That is a context and cache design change, not a second rate cut — do not re-price the July 30 table because the window grew.

When 1M context matters: full-repo review, long legal packs, or multi-turn agent history that previously required chunking.

When it does not: classification/routing lanes already well under 200K. Sending a 1M window “because it is available” inflates uncached input until breakpoints are set.

First-party pricing math (not a client silhouette)

No anonymized engagement is cited for this specific July 30 cut. The numbers below are reproducible arithmetic on the published Bedrock rate card — one illustrative monthly shape, labeled as math.

Illustrative Luna month: 50M input tokens + 5M output tokens (high-volume classification / routing mix).

Direction: −$70.40 / month (−80%) for that token mix on Luna alone. Same volume on Terra post-cut: (50 × $2.20) + (5 × $13.20) = $176.00 — still ~10× Luna for identical token counts, which is why model choice still dominates the bill.

Reproduce this — Open the Bedrock pricing OpenAI table. Multiply your Cost Explorer / CUR input and output token totals by the Luna and Terra rows. For cache-heavy agent loops, include cache-write and cache-read lines separately (90% discount on cache read). If totals diverge, you are missing a service tier or Mantle usage type — see AWS’s CUR Bedrock token-type guide.

Pair this with the broader playbook in Bedrock cost optimization: token budgets and model selection.

Opinionated routing after the cut

We recommend: default new high-QPS classification / extraction / thin-tool lanes to Luna on Bedrock Mantle after a frozen-prompt quality bakeoff. Keep everyday agentic production on Terra until Luna’s pass rate matches. Keep deep overnight / exploit-class / multi-hundred-step agents on Sol (or Claude Opus 5 where ZDR-default + Guardrails composition matter more — see Opus 5 on Bedrock).

Trade-off you accept: lower $ / MTok on Luna vs higher risk of silent quality regressions on nuanced prompts if you promote by price alone.

For the enterprise Bedrock-vs-direct-API frame (IAM, PrivateLink, commitments), see AWS Bedrock vs OpenAI API.

What broke (pattern): “cheaper tokens, same task cost”)

What broke — Teams that only reprice the model ID and leave prompts, max_output / completion budgets, and tool-loop depth unchanged often see token $ drop while cost per completed task barely moves. Root cause: Luna (and Terra) still bill output heavily relative to input; verbose system prompts and uncapped completions dominate. Detect via CUR: output-token $ share stays >50% of the model line. Fix: cap max output per lane, add prompt-cache breakpoints for sticky prefixes (90% cache-read discount), and measure $ / successful task, not $ / 1K tokens.

On-demand still wins for spiky traffic; if you were about to buy Provisioned Throughput for a Luna-class lane, re-run the break-even after this cut — on-demand vs Provisioned Throughput.

Invoke Luna / Terra on Mantle (Responses API)

Context: OpenAI Python SDK against Bedrock Mantle, region us-east-1, model IDs from the Luna / Terra model cards. Auth via Bedrock API key or SigV4 as documented for your path.

# openai>=1.x; OPENAI_BASE_URL=https://bedrock-mantle.us-east-1.api.aws/openai/v1
from openai import OpenAI

client = OpenAI()  # uses OPENAI_API_KEY + OPENAI_BASE_URL from env

response = client.responses.create(
    model='openai.gpt-5.6-luna',
    input='Classify this support ticket into billing, outage, or how-to. Reply with one label only.',
)

print(response)

Swap openai.gpt-5.6-terra for mid-tier lanes. Regions today: us-east-1, us-east-2, us-west-2 for Luna/Terra; Sol is us-east-1 / us-east-2 only. Geo/global inference IDs are not supported on these Mantle model cards at time of writing — plan capacity inside those Regions.

What to Do This Week

1. Pull last 30 days of Bedrock OpenAI token usage from Cost Explorer / CUR; split by model ID.
2. Recompute monthly $ at the new Luna / Terra rows; leave Sol at $5.50 / $33.
3. Pick one high-volume lane that is not Sol-dependent; freeze 50–100 production prompts; score Luna vs current model on pass rate, p95 latency, and $ / completed task.
4. Enable prompt caching on sticky system + tool prefixes; confirm cache-read lines appear in CUR.
5. Update budgets and anomaly detectors — an 80% Luna drop can look like a “usage cliff” to naive alerts (FinOps Agent / Cost Anomaly context).
6. If you are still on OpenAI direct for regulated data, revisit Bedrock for commitment consolidation and IAM — GenAI on AWS / Amazon Bedrock consulting.

What This Post Doesn’t Cover

• A published FactualMinds latency/quality bakeoff of Luna vs Terra vs Claude Haiku/Sonnet on our prompt packs (no harness linked yet).
• OpenAI Fast-mode / Flex / Batch tier matrices beyond pointing at the OpenAI and Bedrock rate cards.
• Codex-on-Bedrock developer UX (covered in the OpenAI models + Codex post).
• EU / non-US regional availability for GPT-5.6 Sol/Terra/Luna beyond the three US Regions AWS listed.

Use the rate card + your CUR. Promote lane-by-lane.

Related reading

• Amazon Bedrock + OpenAI Models, Codex & Managed Agents
• AWS Bedrock vs OpenAI API: Enterprise Decision Guide
• AWS Bedrock Cost Optimization: Token Budgets and Model Selection
• Bedrock Provisioned Throughput vs On-Demand Break-Even
• Claude Opus 5 on AWS Bedrock (July 2026)
• Amazon Bedrock Nova Models Guide
• Generative AI on AWS · Amazon Bedrock services · AWS cost optimization

Frequently asked questions