Flare, Sunburst, and WaveSpeedAI: routing bulk images in Bubble without losing consistency

Why this matters now

Recent digest coverage landed ChatGPT Images 2.5 (gpt-image-2.5-flare, gpt-image-2.5-sunburst), scale of >3 billion images per week (ChatGPT plus GPT-Image API combined—not API-only volume), and WaveSpeedAI with published per-image tables. That sits on top of async/batch and latency work already in your backlog. The new question is multi-vendor routing and pricing tiers, not PNG ratios. Worth reading next: Batch API up to 50k. Same topic: Bubble.io pricing in 2026: which. Worth reading next: Batch API up to 50k.

Flare vs Sunburst on the primary provider (OpenAI)

From public launch materials (2026-09-08):

• Flare — default for everyday / high-volume generation; OpenAI cites up to ~50% lower latency vs Images 2.0.
• Sunburst — precision-heavy creative and editing; longer generation time.

Flare and Sunburst share identical token list pricing ($5/M text in, $8/M image in, $30/M image out on official pricing). Choose speed vs edit fidelity, not a cheaper model ID.

For Bubble apps:

• Route bulk text-to-image to Flare by default.
• Reserve Sunburst for labeled hero jobs, hard edits, or when Flare fails your acceptance tests—not for every bulk slot.

WaveSpeedAI: when price tables matter

WaveSpeed is a third-party host with four REST paths (Flare/Sunburst × text-to-image/edit), an async predict → poll flow, different parameter names (resolution, aspect_ratio vs OpenAI size, quality), and fixed per-image pricing (e.g. text-to-image medium at 1K at $0.024; Flare and Sunburst match on WaveSpeed’s published table).

They explicitly warn: quality tier names are not 1:1 between GPT Image 2 and 2.5—re-baseline on medium, do not assume preset parity.

WaveSpeed fits when:

• You need predictable COGS per image and burst offload.
• A golden-set proves visual parity at mapped settings (often Flare-only for bulk).

Avoid WaveSpeed when:

• You run multi-turn edit chains—keep one vendor and one pinned snapshot.
• You rely on OpenAI Responses continuity—splitting vendors breaks state.

Routing matrix (high-volume Bubble / inferred BIAIG-style connector)

No public BIAIG docs were in the research packet; treat this as inferred from OpenAI/WaveSpeed guidance. Validate on your connector.

Scenario	Route	Rationale
Bulk text-to-image, brand-flex slots	Flare on primary OpenAI	Latency, IPM, single parameter stack
Bulk with fixed $/image COGS	WaveSpeed Flare after golden-set	Table pricing; bake poll into workflow
Multi-step edits, same subject/background	Single vendor, pinned snapshot e.g. gpt-image-2.5-flare-2026-09-08	Same quality label ≠ same visuals across models; test full chains
Hero / precision edit	Sunburst, same vendor as follow-up steps	Do not alternate Flare → Sunburst → another host without tests
Batch economics	gpt-image-2 Batch in an isolated lane if quality passes	Different generation—consistency trade-off
2.5 Batch half-rate	Verify before production	Model pages list Batch for 2.5, but public batch pricing/acceptance for 2.5 IDs was unclear in Sept 2026 coverage—do not budget “50% off Flare” without an official batch row

Consistency checklist (before production routing)

1. Pin model IDs (dated snapshots), not floating “latest” in prod.
2. Fix prompt, references, dimensions, and quality; log vendor + endpoint per edit step in Bubble.
3. Never swap vendor or model mid-chain; use compositing when regions must stay pixel-identical (per OpenAI image prompting guidance).
4. Maintain an explicit OpenAI ↔ WaveSpeed parameter map in the API Connector; document it internally.
5. Run your golden-set on primary vs WaveSpeed at mapped tiers (start at medium).
6. IPM/TPM may force queues or offload—routing is not price-only; do not mix vendors inside one edit thread.
7. OpenAI org verification for GPT Image; review third-party retention, region, and policy separately. Worth reading next: OpenAI Decisions API (preview): classification. A closer look: bubble.io reviews for founders: a.

Open questions (do not fake answers in COGS models)

• Whether /v1/batch accepts 2.5 flare/sunburst with documented half-rate pricing.
• Whether WaveSpeed runs equivalent weights/snapshots and moderation vs direct OpenAI.
• Break-even between OpenAI token billing and WaveSpeed $/image for your quality/resolution mix—requires your calculator and measurements.

Bottom line

Flare is the default bulk lane on primary OpenAI; Sunburst is a narrow premium for edit-heavy work; WaveSpeed is optional predictable COGS after parity proof—not a drop-in for multi-turn chains. Consistency beats a cent per image—design vendor boundaries as carefully as you design async and batch.