GPT-6 Astra & Image 2.5: Official vs Relay API Prices

A new model launches, and discounted API offers appear almost immediately. GPT-6 Astra already has relay quotes near one yuan per million input tokens. Newly released GPT Image 2.5 Flare has a merchant listing at one Chinese cent per call. The difficult part is knowing what the funded account will actually deliver.

We checked GPT-6 Astra, released September 3, and GPT Image 2.5 Flare/Sunburst, released September 8. Astra compares selected SubRouter, OhMyGPT and OpenRouter routes with official standard access. The image section separately lists confirmed SubRouter merchant offers; missing prices are not filled with guesses.

The three platforms' Astra offers and added image screenshots were checked September 10, 2026; the official baseline was checked September 9. Image rows carry their own check dates. SubRouter retains original CNY. Official, OhMyGPT and OpenRouter USD quotes use the site's display conversion of CNY 7.25/USD, not live FX or CNY settlement. No paid inference tests were performed. Commercial disclosure: SubRouter is a ComputeUnion commercial partner, and this article contains a referral link. This is not an endorsement of model identity, price accuracy or uptime.

GPT-6 Astra: a substantial advertised price gap

The official model page lists Standard input/output at USD 10/50 per million tokens—CNY 72.50/362.50 at the display conversion. We do not mix Standard with Batch, Flex or Fast rates. Above 272K input tokens per request, the full request incurs double input rates and 1.5 times output rates.

SubRouter has the lowest selected quote in this table, not an established market-wide minimum. This targeted comparison selects a low-price SubRouter route and higher-priced standard routes on OhMyGPT and OpenRouter. Lower quotes exist elsewhere; consult the full Astra provider list. Other providers remain in the full comparison.

OhMyGPT's published Azure Standard short-context rate already includes its 10% discount. OpenRouter uses OpenAI/Azure standard routes, not its more expensive Fast route. Both list cheaper Flex options; OhMyGPT also lists Batch. These have different processing and timing conditions, so standard rates are not each platform's minimum across all plans. SubRouter is the selected route's public quote, not a universal platform rate; equivalent upstream identity and service have not been established.

This SubRouter route also lists cache reads at CNY 0.18 per million tokens, with cache creation at CNY 2.25 for the five-minute tier and CNY 3.60 for the one-hour tier. The calculation below excludes caching; workloads reusing long prompts should check it separately.

These sources advertise these prices; we have not established equivalent model identity and service. The selected SubRouter route still lists context specifications as unknown, so confirm long-input billing. Its review label describes availability review, not our model-identity or long-term uptime test. We lack sufficient evidence to rank these platforms by reliability.

Reconcile a small workload instead of guessing monthly usage

Suppose several requests accumulate 1M input and 0.2M billable output tokens, without caching or extra tools. This is price arithmetic, not measured performance or anyone's monthly forecast.

For these token counts and selected rates, SubRouter totals CNY 3.60, OhMyGPT CNY 130.50, and OpenRouter CNY 145.00. That is SubRouter's price attraction in this sample, not measured task-level savings. Funding fees, payment FX and extra tools are excluded; usable completion still determines whether the advertised saving translates into a lower bill. OpenRouter's platform fees are listed on its pricing page.

The input total is distributed across requests within the selected tier. SubRouter's sampled context tier is unspecified: its calculation uses the advertised rate only, not a promise for arbitrary prompt lengths. Holding token counts constant isolates prices, not output quality, length or tool usage. Reconcile input, output, cache charges, request records and balance deductions. With retries, compare actual batch charges / accepted tasks.

GPT Image 2 and 2.5: show the cost per call

Image buyers want to know what a product shot and its revisions will cost. The official release positions Flare for speed and everyday creation, and Sunburst for more precise generation and editing. Both have API access. We did not benchmark quality or latency.

Swipe across the table to see all billing conditions.

Each row is a separate route within SubRouter. Platform-name and price links open registration; original offer addresses are retained in the source records. At the listed rates, 100 billed calls cost CNY 1.20 for the 0.012 GPT Image 2 offer, CNY 1.00 for the 0.01 Flare offer, or CNY 8.00 for the 0.08 Sunburst offer. This is not a promise of 100 accepted images. Confirm quantity, specifications and failed-request billing.

The unsuffixed gpt-image-2.5 ID does not establish Flare or Sunburst identity. Two such offers list CNY 0.01 and 0.08 per call; their exact versions remain unspecified, so their raw IDs are retained separately. GPT Image 2 is a different model: inspect its current providers and specifications rather than treating it as a matched-spec 2.5 discount.

Merchant price screenshots from SubRouter

Source: GPT model listings by merchants on the SubRouter platform. User-provided originals, with page prices cross-checked on September 10, 2026; prices and status indicators are unedited. Open a screenshot at full size. The displayed short-window availability is not our long-term reliability test.

There is no invented universal official per-image price here. Official billing depends on the request, and the documentation says the GPT Image 2 calculator does not estimate 2.5 consumption. Comparing an older model's low-quality small image with the new model's high-quality large image would manufacture a meaningless discount.

Fix a specification you need: for example, one 1024×1024 product image at a named quality, using the same reference and preserving label text while replacing its background. This is a proposed test, not verified support by these merchants. Ask whether the displayed price covers it. Unspecified dimensions, quality or quantity do not imply the highest tier.

A few cents saved may disappear if a route requires extra edits. For bulk social assets, an acceptable low per-call rate can be attractive. Preserve the product and text first, assess speed next, then compare cost per accepted image. The Flare records and Sunburst records may include both per-call and usage-based offers; keep those units separate.

Ask what the multiplier multiplies

Some pages show base rates before a group adjustment; others show final CNY. Some use dollar-labelled credits with their own cash conversion. A hypothetical 0.10-credit call costs CNY 0.10 if one yuan buys one credit, but CNY 0.70 if seven yuan buys it. This is not a sampled provider's stated top-up policy.

Confirm cash conversion, selected group and final displayed price, then reconcile a small deduction. Do not apply discounts or FX again to a final CNY quote. Some SubRouter API fields were normalized, so we checked actual merchant pages. SubRouter's 0.0248× means approximately 2.48% of official input pricing, not a promised discount across all features. Cache, context tiers, tools, image settings and failures have separate rules.

SubRouter's useful difference is not only its cheapest quote

Multiple models mean scattered keys, balances and fallback endpoints. SubRouter's marketplace gathers merchant prices, interfaces and public service information, with fixed-merchant selection. Its documentation describes unified access; its routing guide describes eligible-provider failover and request/billing records. Original addresses are retained in the source records. We confirmed these visible features and descriptions, not actual failover behavior.

That is useful when you need coding and image models plus backup merchants. Begin with a fixed route so each test has an identifiable supplier. Check backup features, prices, privacy conditions and caching before enabling smart routing. Automatic switching does not make every fallback suitable for the same task.

OhMyGPT identifies Azure sourcing and multiple billing tiers. OpenRouter exposes providers, route prices and routing information. Those are useful selection criteria, not advantages erased by a higher listed rate. Frequent low-price model switching makes SubRouter worth testing; requirements around sourcing, interfaces and service tiers should still be checked alongside price.

Spend a little to discover expensive problems first

Use 10–20 real tasks, not one hello. Test code tools and iterative edits, or fixed product references and delivery dimensions. Hold model, merchant, group and parameters constant; record completion time, errors, acceptance and actual charges. Repeat during another normal working period.

A small sample can expose incompatibility or misunderstood billing. It cannot establish long-term uptime or identify the upstream model from a returned model string. We assign no reliability scores without supporting evidence.

• Route: model ID, merchant, group, automatic switching and fallback price.
• Bill: top-up conversion, tier, failures/retries, images per call and exportable request records.
• Delivery: required code tools or image dimensions, quality, references and edit support, tested with your work.
• Support: refunds, contacts, invoice issuer, retention and data-use terms. Domain control does not replace service commitments.

Low prices do not prove a model problem; high prices do not prove suitability. Keep official access as a candidate when specific tools, contracts or data arrangements matter. Useful savings mean the task still gets done and the actual bill falls—not discovering the restrictions after a large deposit.

This is a dated snapshot, not a live feed. Recheck Astra's current records and the image API directory before buying. Models change; comparing currency, route, billing unit and accepted output remains useful.

Sources and scope

Release dates and positioning come from the Astra announcement and linked Images 2.5 release. Astra rates come from its model page. Image boundaries were checked in the Flare and Sunburst documentation. The tables link OhMyGPT and OpenRouter sources; SubRouter's original offer and documentation addresses are recorded below. USD rates use the same conversion and faster premium tiers are not substituted for standard pricing. See the collection methodology.

Original SubRouter quote and documentation addresses

These original source addresses can be copied to inspect the records. SubRouter platform names, prices and visit buttons open registration; IDs in the table match the quote addresses below.

• GPT-6 Astrahttps://subrouter.ai/models/11029
• gpt-image-2 · ¥0.012 / callhttps://subrouter.ai/models/1989
• gpt-image-2.5-flare · ¥0.01 / callhttps://subrouter.ai/models/11436
• gpt-image-2.5-flare · ¥0.07 / callhttps://subrouter.ai/models/11402
• gpt-image-2.5-flare · ¥0.17 / callhttps://subrouter.ai/models/11434
• gpt-image-2.5-sunburst · ¥0.08 / callhttps://subrouter.ai/models/11401
• gpt-image-2.5-sunburst · ¥0.17 / callhttps://subrouter.ai/models/11435
• gpt-image-2.5 · ¥0.01 / callhttps://subrouter.ai/models/11395
• gpt-image-2.5 · ¥0.08 / callhttps://subrouter.ai/models/11386
• Marketplacehttps://subrouter.ai/models
• Documentationhttps://subrouter.ai/docs
• Routing guidehttps://subrouter.ai/openrouter-alternative

This targeted three-platform sample is not a market-wide ranking. All image offers are routes within SubRouter, not separate aggregators. We did not perform paid inference, matched-specification image tests, refunds or long-term monitoring.