Bedrock Cuts GPT-5.6 Prices by up to 80%: What the Luna/Terra/Sol Asymmetry Tells You

On July 30, AWS announced up to 80% lower prices for OpenAI GPT-5.6 models on Amazon Bedrock. Price-cut announcements usually earn a shrug and an updated spreadsheet cell. This one deserves ten minutes of your attention, because the shape of the cut — which tiers moved, by how much, and which didn’t move at all — says a lot about where inference economics are heading, and it materially changes the routing math for anyone running GPT-5.6 workloads on Bedrock.

Here are the numbers first, then the analysis.

The new price sheet

Model	Change	Input ($/1M tokens)	Output ($/1M tokens)
GPT-5.6 Luna	−80%	$0.22	$1.32
GPT-5.6 Terra	−20%	$2.20	$13.20
GPT-5.6 Sol	unchanged	$5.50	$33.00

Three details worth pinning:

• Cache reads on Luna are $0.022 per 1M tokens — 10% of the input price. Keep that number in mind; it compounds with the cut in a way I’ll quantify below.
• The change applied automatically on July 30. No new SKU, no opt-in, no customer action.
• It’s live in three US regions only: US East (N. Virginia), US East (Ohio), and US West (Oregon). More on that constraint later, because it’s the biggest one.

The GPT-5.6 family went GA on Bedrock on July 13. The price cut landed seventeen days later — the same day OpenAI cut its own first-party prices for the same models. That timing is not a coincidence, and it’s where the analysis starts.

Why this cut happened (and why same-day)

The inference price war has a new rule: the aggregator matches the source, immediately.

Bedrock’s pitch for hosting OpenAI models was never “cheaper than OpenAI.” It’s “the same models inside your AWS boundary”: IAM for auth, VPC endpoints for network isolation, CloudWatch for observability, Guardrails for policy, one bill for everything. That pitch collapses the moment a price gap opens — a CFO will tolerate a governance premium of zero percent, roughly. So when OpenAI cut Luna and Terra first-party, Bedrock had to match within the news cycle, and did.

For us as architects, the strategic read is simple: AWS wants zero price-based reasons to route around Bedrock. The margin play is the platform (and the GPUs it sells to everyone else), not a markup on tokens. Expect this pattern to repeat: any first-party cut by a model provider hosted on Bedrock should propagate to Bedrock within days. Factor that into build-vs-wait decisions — “it’s cheaper direct” is now a transient state, not an architecture input.

The asymmetry is the message

Look at the cut again: −80% on the small tier, −20% on the mid tier, 0% on the frontier tier. If this were just margin compression, you’d expect a uniform trim. It isn’t. It’s a statement about where price elasticity actually lives.

• Luna (−80%) is the volume tier — AWS positions it for content processing, classification, and customer-service automation. Demand here is brutally elastic: at $1.10/1M input nobody classifies their entire email archive; at $0.22/1M input, “run the model over everything” becomes a line item you can approve without a meeting. The cut isn’t generosity — it’s an attempt to unlock workload classes that didn’t exist at the old price.
• Terra (−20%) is the balanced production tier for sophisticated reasoning. Demand is moderately elastic — real workloads live here, but they were mostly viable already. The trim keeps it honest against competitors’ mid-tiers.
• Sol (unchanged) is the frontier. If you need frontier reasoning, you’re not price-shopping — you’re capability-shopping. Zero discount is exactly what pricing power looks like.

The resulting input-price ladder is 1 : 10 : 25 (Luna : Terra : Sol). Before the cut it was 1 : 2.5 : 5. That’s a fundamentally different routing landscape: the penalty for sending a request one tier too high used to be 2.5×; it’s now 10×.

What it changes in your routing math

If you run a model router (or have been putting off building one), this cut just changed your payback period. A concrete blend: a router that resolves 80% of traffic to Luna and 20% to Terra now pays a blended input price of

0.8 × $0.22 + 0.2 × $2.20 = $0.62 per 1M tokens

versus $2.20 if you lazily send everything to Terra. That’s −72% on input spend for a two-line routing policy. Before the cut, the same blend saved you 55% — meaningful, but not “cancel the meeting” meaningful. At the new ladder, tier discipline is one of the highest-leverage FinOps actions available on a GenAI stack.

Worked example 1: customer-service assistant on Luna

Take a support assistant handling 50,000 conversations/day, 6 turns each, ~3,000 input tokens (system prompt + history + context) and ~250 output tokens per turn. That’s 900M input and 75M output tokens per day.

	Input/day	Output/day	Total/day	~Monthly
Before (at $1.10 / $6.60)	$990	$495	$1,485	~$44,600
After (at $0.22 / $1.32)	$198	$99	$297	~$8,900
After + 60% cache-hit on input	$91	$99	$190	~$5,700

The last row is the one to internalize. Support conversations re-send the same system prompt and history every turn — prime caching territory. At $0.022/1M for cache reads, a 60% cache-hit rate takes the daily input bill from $198 to $91. Stacked with the price cut, the workload lands ~87% below its July bill. If you evaluated a “classify and draft-respond to everything” project six months ago and killed it on cost, that evaluation is stale.

Worked example 2: agentic pipeline on Terra

An agentic document-processing pipeline: 2,000 runs/day, 15 model calls per run, ~8,000 input + ~1,200 output tokens per call. That’s 240M input, 36M output per day.

	Total/day	~Monthly
Before (at $2.75 / $16.50)	$1,254	~$37,600
After (at $2.20 / $13.20)	$1,003	~$30,100

A clean −20%, ~$7,500/month back, for doing nothing. Nice, but notice the gap to the Luna example: the real money is not in the Terra discount — it’s in auditing those 15 calls per run and asking how many of them actually need Terra. If 10 of the 15 are extraction and formatting steps that Luna handles, the blended cost drops by another ~60% on those calls. The cut doesn’t just lower prices; it raises the ROI of auditing your call graph.

The limits (read before migrating)

Four constraints keep this from being a pure win, and the first one is the big one:

1. US-only. N. Virginia, Ohio, Oregon. If your workload lives in Europe under data-residency constraints, these prices are not for you (yet) — cross-region inference into US regions is exactly what your residency review exists to prevent. EU-based teams should treat this as a preview of pricing that may eventually reach European regions, not as an action item.
2. Different API path. GPT-5.6 models on Bedrock are consumed through the OpenAI Responses API on the bedrock-mantle endpoint, not the native Converse API. Porting an existing Converse-based stack is a client change, not a config change — and you should verify feature parity for anything you lean on (Guardrails integration, structured output, tool-calling behavior) before moving volume.
3. Quotas. An 80% price cut is an invitation to 5× your traffic; your TPM/RPM quotas did not get the same multiplier. If the new economics make a big-volume workload viable, put the quota-increase request in before the migration, not during the incident.
4. Long-context pricing. A 1M-token long-context tier exists for GPT-5.6, but the announcement doesn’t state its rates. Don’t assume base pricing applies to long-context calls — verify on the Bedrock pricing page before designing around it.

What I’d actually do this week

• Running GPT-5.6 on Bedrock in a US region? Nothing to migrate — the cut already hit your bill on July 30. Check Cost Explorer, confirm the drop, and bank the delta.
• Running everything on Terra or Sol? Audit the call graph. The 10× Luna-Terra gap makes a classification-based router (or even a static per-step tier assignment) worth building this sprint, not this quarter.
• Not caching prompts? On Luna, cache reads at 10% of an already −80% input price is the single cheapest token source on the platform. Restructure prompts for prefix stability (static system prompt first, volatile context last) and measure your hit rate.
• Killed a high-volume GenAI project on cost in H1? Re-run the numbers. An 80% cut plus caching moves most “too expensive at scale” verdicts.

The price war between model providers has fully arrived inside the aggregators, and Bedrock just demonstrated its response latency is measured in hours. For once, being downstream of a price war is exactly where you want your workloads sitting.

Sources: AWS What’s New — OpenAI GPT-5.6 Terra and Luna pricing update on Amazon Bedrock (July 30, 2026), AWS News Feed article, GPT-5.6 GA announcement (July 13, 2026). Pricing figures as published at announcement time; verify current rates on the Bedrock pricing page. I work as a Solutions Architect at AWS; opinions here are my own.