Model capabilities

Provider output: 128000

Streaming: Unknown

Effort values: Unknown

Provider formats: image/jpeg, image/png, image/gif, image/webp

Bytes per image: Unknown

Aggregate encoded bytes: Unknown

Hard width / height: 8000 / 8000

Direct Claude API: 10 MB base64-encoded per image; 32 MB whole request. MB byte definition unspecified.

Animation uses the first frame. More than 20 images has stricter dimensions; Pumpkin caps requests at 20 images.

Pumpkin-selected image limits, not provider maxima:

• Pixel area: 3750656
• Decoded allocation: 100000000 bytes
• File bytes per image: 7000000 (Pumpkin cap; 7,000,000 file bytes encode to at most 9,333,336 base64 bytes, below the documented 10 MB)
• Aggregate encoded image data: 20000000 bytes
• Output format: image/png
• 28 px patches, budget 4784 per image

Preserve aspect ratio; never upscale. Static GIF/WebP can be normalized; animated GIF is rejected. Provider-listed BMP/TIFF/PPM does not imply Pumpkin decoder support. Unsupported inputs fail explicitly.

Editorial guidance: Hardest work: architecture, multi-service changes, long-context debugging, anything where a wrong call is expensive.

Standard Messages output limit; thinking shares output budget.

Adaptive thinking is always on; forced tool choice is unsupported.

Verified Sep 22, 2026

Provider output: 128000

Streaming: Unknown

Effort values: Unknown

Provider formats: image/jpeg, image/png, image/gif, image/webp

Bytes per image: Unknown

Aggregate encoded bytes: Unknown

Hard width / height: 8000 / 8000

Direct Claude API: 10 MB base64-encoded per image; 32 MB whole request. MB byte definition unspecified.

Animation uses the first frame. More than 20 images has stricter dimensions; Pumpkin caps requests at 20 images.

Pumpkin-selected image limits, not provider maxima:

• Pixel area: 3750656
• Decoded allocation: 100000000 bytes
• File bytes per image: 7000000 (Pumpkin cap; 7,000,000 file bytes encode to at most 9,333,336 base64 bytes, below the documented 10 MB)
• Aggregate encoded image data: 20000000 bytes
• Output format: image/png
• 28 px patches, budget 4784 per image

Preserve aspect ratio; never upscale. Static GIF/WebP can be normalized; animated GIF is rejected. Provider-listed BMP/TIFF/PPM does not imply Pumpkin decoder support. Unsupported inputs fail explicitly.

Editorial guidance: Long-running agentic coding and knowledge work. Newer and cheaper than Opus 5, still below Fable.

Standard Messages output limit; thinking shares output budget.

Adaptive thinking is always on; forced tool choice is unsupported. Default effort is medium.

Cache read is 0.05x base input, not the usual 0.1x.

Verified Sep 22, 2026

Provider output: 128000

Streaming: Unknown

Effort values: Unknown

Provider formats: image/jpeg, image/png, image/gif, image/webp

Bytes per image: Unknown

Aggregate encoded bytes: Unknown

Hard width / height: 8000 / 8000

Direct Claude API: 10 MB base64-encoded per image; 32 MB whole request. MB byte definition unspecified.

Animation uses the first frame. More than 20 images has stricter dimensions; Pumpkin caps requests at 20 images.

Pumpkin-selected image limits, not provider maxima:

• Pixel area: 3750656
• Decoded allocation: 100000000 bytes
• File bytes per image: 7000000 (Pumpkin cap; 7,000,000 file bytes encode to at most 9,333,336 base64 bytes, below the documented 10 MB)
• Aggregate encoded image data: 20000000 bytes
• Output format: image/png
• 28 px patches, budget 4784 per image

Preserve aspect ratio; never upscale. Static GIF/WebP can be normalized; animated GIF is rejected. Provider-listed BMP/TIFF/PPM does not imply Pumpkin decoder support. Unsupported inputs fail explicitly.

Editorial guidance: Strong implementation and review when Fable is overkill: scoped features, correctness-sensitive edits, production-ready diffs.

Standard Messages output limit; thinking shares output budget.

Verified Sep 22, 2026

Provider output: 128000

Streaming: Unknown

Effort values: Unknown

Provider formats: image/jpeg, image/png, image/gif, image/webp

Bytes per image: Unknown

Aggregate encoded bytes: Unknown

Hard width / height: 8000 / 8000

Direct Claude API: 10 MB base64-encoded per image; 32 MB whole request. MB byte definition unspecified.

Animation uses the first frame. More than 20 images has stricter dimensions; Pumpkin caps requests at 20 images.

Pumpkin-selected image limits, not provider maxima:

• Pixel area: 3750656
• Decoded allocation: 100000000 bytes
• File bytes per image: 7000000 (Pumpkin cap; 7,000,000 file bytes encode to at most 9,333,336 base64 bytes, below the documented 10 MB)
• Aggregate encoded image data: 20000000 bytes
• Output format: image/png
• 28 px patches, budget 4784 per image

Preserve aspect ratio; never upscale. Static GIF/WebP can be normalized; animated GIF is rejected. Provider-listed BMP/TIFF/PPM does not imply Pumpkin decoder support. Unsupported inputs fail explicitly.

Editorial guidance: Everyday coding: routine features, refactors with a known shape, tests, docs. Mid-tier default for Anthropic.

Standard Messages output limit; thinking shares output budget.

Verified Sep 22, 2026

Provider output: 64000

Streaming: Unknown

Effort values: Unknown

Provider formats: image/jpeg, image/png, image/gif, image/webp

Bytes per image: Unknown

Aggregate encoded bytes: Unknown

Hard width / height: 8000 / 8000

Direct Claude API: 10 MB base64-encoded per image; 32 MB whole request. MB byte definition unspecified.

Animation uses the first frame. More than 20 images has stricter dimensions; Pumpkin caps requests at 20 images.

Pumpkin-selected image limits, not provider maxima:

• Pixel area: 1229312
• Decoded allocation: 100000000 bytes
• File bytes per image: 7000000 (Pumpkin cap; 7,000,000 file bytes encode to at most 9,333,336 base64 bytes, below the documented 10 MB)
• Aggregate encoded image data: 20000000 bytes
• Output format: image/png
• 28 px patches, budget 1568 per image

Preserve aspect ratio; never upscale. Static GIF/WebP can be normalized; animated GIF is rejected. Provider-listed BMP/TIFF/PPM does not imply Pumpkin decoder support. Unsupported inputs fail explicitly.

Editorial guidance: Cheap, short jobs: mechanical edits, triage, titles, compaction, apply this exact change. Not for design or long hunts.

Standard Messages output limit; thinking shares output budget.

Verified Sep 22, 2026

Provider output: Unknown separate numeric maximum

Streaming: Unknown

Effort values: low, medium, high, xhigh

Provider formats: image/jpeg, image/png

Bytes per image: Unknown

Aggregate encoded bytes: Unknown

Hard width / height: Unknown / Unknown

20 MiB per image; decoded versus encoded accounting unspecified. No image-count limit stated, but context limits still apply.

Hard pixel dimensions and numeric provider resize recommendation unknown.

Pumpkin-selected image limits, not provider maxima:

• Pixel area: 2500000
• Decoded allocation: 100000000 bytes
• Aggregate encoded image data: 20000000 bytes
• Output format: image/png

Preserve aspect ratio; never upscale. Static GIF/WebP can be normalized; animated GIF is rejected. Provider-listed BMP/TIFF/PPM does not imply Pumpkin decoder support. Unsupported inputs fail explicitly.

Editorial guidance: xAI flagship for coding and agent work. Same price as 4.6, stronger on tools and harder implementation.

Provider documents no separate text output limit. maxOutput=128000 is Pumpkin request policy, not a provider maximum.

Reasoning cannot be disabled.

Verified Sep 22, 2026

Provider output: Unknown separate numeric maximum

Streaming: Unknown

Effort values: low, medium, high, xhigh

Provider formats: image/jpeg, image/png

Bytes per image: Unknown

Aggregate encoded bytes: Unknown

Hard width / height: Unknown / Unknown

20 MiB per image; decoded versus encoded accounting unspecified. No image-count limit stated, but context limits still apply.

Hard pixel dimensions and numeric provider resize recommendation unknown.

Pumpkin-selected image limits, not provider maxima:

• Pixel area: 2500000
• Decoded allocation: 100000000 bytes
• Aggregate encoded image data: 20000000 bytes
• Output format: image/png

Preserve aspect ratio; never upscale. Static GIF/WebP can be normalized; animated GIF is rejected. Provider-listed BMP/TIFF/PPM does not imply Pumpkin decoder support. Unsupported inputs fail explicitly.

Editorial guidance: Fast workhorse: investigation, evidence collection, intermediate code, image analysis. Good value; less stubborn than Fable on long stuck turns.

Provider documents no separate text output limit. maxOutput=128000 is Pumpkin request policy, not a provider maximum.

Reasoning cannot be disabled.

Verified Sep 22, 2026

Provider output: 128000

Streaming: Yes

Effort values: none, low, medium, high, xhigh, max

Provider formats: image/png, image/jpeg, image/webp, image/gif

Bytes per image: Unknown

Aggregate encoded bytes: Unknown

Hard width / height: Unknown / Unknown

GIF must be nonanimated. Total request payload up to 512 MB; byte accounting unspecified.

Original/auto preserves resolution subject to a 65535px preprocessing cap; more than 30000 processed 32px patches per image is rejected. Pumpkin chooses a smaller high-detail patch budget, not original-detail fidelity.

Pumpkin-selected image limits, not provider maxima:

• Pixel area: 2560000
• Decoded allocation: 100000000 bytes
• Aggregate encoded image data: 20000000 bytes
• Output format: image/png
• 32 px patches, budget 2500 per image

Preserve aspect ratio; never upscale. Static GIF/WebP can be normalized; animated GIF is rejected. Provider-listed BMP/TIFF/PPM does not imply Pumpkin decoder support. Unsupported inputs fail explicitly.

Editorial guidance: Solid mid OpenAI: focused features, bug fixes with a known cause, normal coding loops.

Maximum input 922000 tokens; reasoning and visible output share the output budget.

Promotional prices at least through 2026-11-21; subsequent rates unknown.

Verified Sep 22, 2026

Provider output: 128000

Streaming: Yes

Effort values: none, low, medium, high, xhigh, max

Provider formats: image/png, image/jpeg, image/webp, image/gif

Bytes per image: Unknown

Aggregate encoded bytes: Unknown

Hard width / height: Unknown / Unknown

GIF must be nonanimated. Total request payload up to 512 MB; byte accounting unspecified.

Original/auto preserves resolution subject to a 65535px preprocessing cap; more than 30000 processed 32px patches per image is rejected. Pumpkin chooses a smaller high-detail patch budget, not original-detail fidelity.

Pumpkin-selected image limits, not provider maxima:

• Pixel area: 2560000
• Decoded allocation: 100000000 bytes
• Aggregate encoded image data: 20000000 bytes
• Output format: image/png
• 32 px patches, budget 2500 per image

Preserve aspect ratio; never upscale. Static GIF/WebP can be normalized; animated GIF is rejected. Provider-listed BMP/TIFF/PPM does not imply Pumpkin decoder support. Unsupported inputs fail explicitly.

Editorial guidance: Small, well-specified work: one-file edits, straightforward tests, low-stakes helpers.

Maximum input 922000 tokens; reasoning and visible output share the output budget.

Verified Sep 22, 2026

Provider output: 128000

Streaming: Yes

Effort values: none, low, medium, high, xhigh, max

Provider formats: image/png, image/jpeg, image/webp, image/gif

Bytes per image: Unknown

Aggregate encoded bytes: Unknown

Hard width / height: Unknown / Unknown

GIF must be nonanimated. Total request payload up to 512 MB; byte accounting unspecified.

Original/auto preserves resolution subject to a 65535px preprocessing cap; more than 30000 processed 32px patches per image is rejected. Pumpkin chooses a smaller high-detail patch budget, not original-detail fidelity.

Pumpkin-selected image limits, not provider maxima:

• Pixel area: 2560000
• Decoded allocation: 100000000 bytes
• Aggregate encoded image data: 20000000 bytes
• Output format: image/png
• 32 px patches, budget 2500 per image

Preserve aspect ratio; never upscale. Static GIF/WebP can be normalized; animated GIF is rejected. Provider-listed BMP/TIFF/PPM does not imply Pumpkin decoder support. Unsupported inputs fail explicitly.

Editorial guidance: Cheapest OpenAI: tiny mechanical tasks, classification, short rewrites. Skip anything that needs judgment.

Maximum input 922000 tokens; reasoning and visible output share the output budget.

Verified Sep 22, 2026

Provider output: 128000

Streaming: Yes

Effort values: low, medium, high, xhigh, max

Provider formats: image/png, image/jpeg, image/webp, image/gif

Bytes per image: Unknown

Aggregate encoded bytes: Unknown

Hard width / height: Unknown / Unknown

GIF must be nonanimated. Total request payload up to 512 MB; byte accounting unspecified.

Original/auto preserves resolution subject to a 65535px preprocessing cap; more than 30000 processed 32px patches per image is rejected. Pumpkin chooses a smaller high-detail patch budget, not original-detail fidelity.

Pumpkin-selected image limits, not provider maxima:

• Pixel area: 2560000
• Decoded allocation: 100000000 bytes
• Aggregate encoded image data: 20000000 bytes
• Output format: image/png
• 32 px patches, budget 2500 per image

Preserve aspect ratio; never upscale. Static GIF/WebP can be normalized; animated GIF is rejected. Provider-listed BMP/TIFF/PPM does not imply Pumpkin decoder support. Unsupported inputs fail explicitly.

Editorial guidance: OpenAI flagship: hard implementation and long context, same class as Fable. Use when the problem is large or ambiguous.

Maximum input 922000 tokens; reasoning and visible output share the output budget.

Tool calling requires Responses; Chat Completions supports text only.

Verified Sep 22, 2026

Provider output: Unknown separate numeric maximum

Pumpkin output request cap: 65536 tokens, not a documented provider maximum.

Streaming: Yes

Effort values: minimal, low, medium, high, xhigh, max

Provider formats: image/png, image/jpeg, image/webp, image/gif

Bytes per image: Unknown

Aggregate encoded bytes: Unknown

Hard width / height: Unknown / Unknown

Images as base64 data URLs or Files API references; audio understanding is degraded on 1.3 (use 1.2), and GIF uses the first frame.

Meta documents no per-image or per-request image byte limits; Pumpkin applies its own conservative client policy rather than trusting an unspecified vendor cap.

Pumpkin-selected image limits, not provider maxima:

• Pixel area: 2500000
• Decoded allocation: 100000000 bytes
• Aggregate encoded image data: 20000000 bytes
• Output format: image/png
• 32 px patches, budget 2500 per image

Preserve aspect ratio; never upscale. Static GIF/WebP can be normalized; animated GIF is rejected. Provider-listed BMP/TIFF/PPM does not imply Pumpkin decoder support. Unsupported inputs fail explicitly.

Editorial guidance: Meta's agentic reasoning model: multi-step tool loops and coding, 1M context, cheap cached input.

Reasoning cannot be turned off (effort "none" returns HTTP 400); reasoning shares the output budget with visible text.

Meta documents no per-request output ceiling; 65536 is a safe request cap below the 1,048,576 context window (input and output share one budget).

Verified Sep 22, 2026

Provider output: Unknown separate numeric maximum

Streaming: Unknown

Effort values: Unknown

Editorial guidance: Strongest DeepSeek: hard coding, agents and long reasoning; Fireworks' top open pick.

Provider output limit Unknown. thinkingMode=none and the 16384 output cap describe Pumpkin adapter policy, not absence of provider reasoning.

Exact context Unknown until metadata discovery; rounded public context labels are not treated as exact limits.

Serverless retirement announced for 2026-09-25; exact cutoff time unknown. Dedicated deployments are excluded.

Verified Sep 22, 2026

Provider output: Unknown separate numeric maximum

Streaming: Unknown

Effort values: Unknown

Provider formats: image/png, image/jpeg, image/gif, image/bmp, image/tiff, image/x-portable-pixmap

Bytes per image: Unknown

Aggregate encoded bytes: Unknown

Hard width / height: Unknown / Unknown

Aggregate base64 image data must be less than 10 MB; MB definition unspecified. URL images less than 5 MB and download within 1.5 seconds.

No numeric universal provider resize recommendation. Pumpkin uses its conservative resize policy.

Pumpkin-selected image limits, not provider maxima:

• Pixel area: 2500000
• Decoded allocation: 100000000 bytes
• Aggregate encoded image data: 9000000 bytes
• Output format: image/png

Preserve aspect ratio; never upscale. Static GIF/WebP can be normalized; animated GIF is rejected. Provider-listed BMP/TIFF/PPM does not imply Pumpkin decoder support. Unsupported inputs fail explicitly.

Editorial guidance: Kimi K3 flagship: hard coding and research agents, image input, 1M context; the priciest here.

Provider output limit Unknown. thinkingMode=none and the 16384 output cap describe Pumpkin adapter policy, not absence of provider reasoning.

Verified Sep 22, 2026

Provider output: Unknown separate numeric maximum

Streaming: Unknown

Effort values: Unknown

Editorial guidance: Newest GLM flagship: coding, reasoning and agent tool use; 1M context.

Provider output limit Unknown. thinkingMode=none and the 16384 output cap describe Pumpkin adapter policy, not absence of provider reasoning.

Exact context Unknown until metadata discovery; rounded public context labels are not treated as exact limits.

Verified Sep 22, 2026

Provider output: Unknown separate numeric maximum

Streaming: Unknown

Effort values: Unknown

Editorial guidance: GLM 5.2 flagship: coding and agent tool use; 1M context.

Provider output limit Unknown. thinkingMode=none and the 16384 output cap describe Pumpkin adapter policy, not absence of provider reasoning.

Exact context Unknown until metadata discovery; rounded public context labels are not treated as exact limits.

Serverless retirement announced for 2026-09-25; exact cutoff time unknown. Dedicated deployments are excluded.

Verified Sep 22, 2026

Provider output: Unknown separate numeric maximum

Streaming: Unknown

Effort values: Unknown

Provider formats: image/png, image/jpeg, image/gif, image/bmp, image/tiff, image/x-portable-pixmap

Bytes per image: Unknown

Aggregate encoded bytes: Unknown

Hard width / height: Unknown / Unknown

Aggregate base64 image data must be less than 10 MB; MB definition unspecified. URL images less than 5 MB and download within 1.5 seconds.

No numeric universal provider resize recommendation. Pumpkin uses its conservative resize policy.

Pumpkin-selected image limits, not provider maxima:

• Pixel area: 2500000
• Decoded allocation: 100000000 bytes
• Aggregate encoded image data: 9000000 bytes
• Output format: image/png

Preserve aspect ratio; never upscale. Static GIF/WebP can be normalized; animated GIF is rejected. Provider-listed BMP/TIFF/PPM does not imply Pumpkin decoder support. Unsupported inputs fail explicitly.

Editorial guidance: Kimi K2.7 Code: coding-tuned agent model with image input; 262K context.

Provider output limit Unknown. thinkingMode=none and the 16384 output cap describe Pumpkin adapter policy, not absence of provider reasoning.

Exact context Unknown until metadata discovery; rounded public context labels are not treated as exact limits.

Serverless retirement announced for 2026-09-25; exact cutoff time unknown. Dedicated deployments are excluded.

Verified Sep 22, 2026

Provider output: Unknown separate numeric maximum

Streaming: Unknown

Effort values: Unknown

Provider formats: image/png, image/jpeg, image/gif, image/bmp, image/tiff, image/x-portable-pixmap

Bytes per image: Unknown

Aggregate encoded bytes: Unknown

Hard width / height: Unknown / Unknown

Aggregate base64 image data must be less than 10 MB; MB definition unspecified. URL images less than 5 MB and download within 1.5 seconds.

No numeric universal provider resize recommendation. Pumpkin uses its conservative resize policy.

Pumpkin-selected image limits, not provider maxima:

• Pixel area: 2500000
• Decoded allocation: 100000000 bytes
• Aggregate encoded image data: 9000000 bytes
• Output format: image/png

Preserve aspect ratio; never upscale. Static GIF/WebP can be normalized; animated GIF is rejected. Provider-listed BMP/TIFF/PPM does not imply Pumpkin decoder support. Unsupported inputs fail explicitly.

Editorial guidance: Kimi K2.6: agents with tool use, images and documents; 262K context.

Provider output limit Unknown. thinkingMode=none and the 16384 output cap describe Pumpkin adapter policy, not absence of provider reasoning.

Exact context Unknown until metadata discovery; rounded public context labels are not treated as exact limits.

Serverless retirement announced for 2026-09-25; exact cutoff time unknown. Dedicated deployments are excluded.

Verified Sep 22, 2026

Provider output: Unknown separate numeric maximum

Streaming: Unknown

Effort values: Unknown

Editorial guidance: Qwen 3.8 (2.4T MoE, 95B active): strong general coding and reasoning; 262K context.

Provider output limit Unknown. thinkingMode=none and the 16384 output cap describe Pumpkin adapter policy, not absence of provider reasoning.

Exact context Unknown until metadata discovery; rounded public context labels are not treated as exact limits.

Public model page says serverless not supported. No substitution of qwen3p8-max pricing.

Verified Sep 22, 2026

Provider output: Unknown separate numeric maximum

Streaming: Unknown

Effort values: Unknown

Editorial guidance: MiniMax M3: fast coding and agent model, good value; 512K context.

Provider output limit Unknown. thinkingMode=none and the 16384 output cap describe Pumpkin adapter policy, not absence of provider reasoning.

Exact context Unknown until metadata discovery; rounded public context labels are not treated as exact limits.

Exact serving-model feature metadata says no images despite generic multimodal architecture prose.

Verified Sep 22, 2026

Provider output: Unknown separate numeric maximum

Streaming: Unknown

Effort values: Unknown

Provider formats: image/png, image/jpeg, image/gif, image/bmp, image/tiff, image/x-portable-pixmap

Bytes per image: Unknown

Aggregate encoded bytes: Unknown

Hard width / height: Unknown / Unknown

Aggregate base64 image data must be less than 10 MB; MB definition unspecified. URL images less than 5 MB and download within 1.5 seconds.

No numeric universal provider resize recommendation. Pumpkin uses its conservative resize policy.

Pumpkin-selected image limits, not provider maxima:

• Pixel area: 2500000
• Decoded allocation: 100000000 bytes
• Aggregate encoded image data: 9000000 bytes
• Output format: image/png

Preserve aspect ratio; never upscale. Static GIF/WebP can be normalized; animated GIF is rejected. Provider-listed BMP/TIFF/PPM does not imply Pumpkin decoder support. Unsupported inputs fail explicitly.

Editorial guidance: DeepSeek V4.1 Flash with image input: quick coding and agent steps at low cost; 1M context.

Provider output limit Unknown. thinkingMode=none and the 16384 output cap describe Pumpkin adapter policy, not absence of provider reasoning.

Exact context Unknown until metadata discovery; rounded public context labels are not treated as exact limits.

Verified Sep 22, 2026

Provider output: Unknown separate numeric maximum

Streaming: Unknown

Effort values: Unknown

Editorial guidance: NVIDIA Nemotron 3 Ultra (preview): large reasoning model; 262K context.

Provider output limit Unknown. thinkingMode=none and the 16384 output cap describe Pumpkin adapter policy, not absence of provider reasoning.

Exact context Unknown until metadata discovery; rounded public context labels are not treated as exact limits.

Verified Sep 22, 2026

Provider output: Unknown separate numeric maximum

Streaming: Unknown

Effort values: Unknown

Editorial guidance: Fast, cheap DeepSeek: extraction, summaries and everyday agent steps; 1M context.

Provider output limit Unknown. thinkingMode=none and the 16384 output cap describe Pumpkin adapter policy, not absence of provider reasoning.

Exact context Unknown until metadata discovery; rounded public context labels are not treated as exact limits.

Serverless retirement announced for 2026-09-25; exact cutoff time unknown. Dedicated deployments are excluded.

Verified Sep 22, 2026

Provider output: Unknown separate numeric maximum

Streaming: Unknown

Effort values: Unknown

Provider formats: image/png, image/jpeg, image/gif, image/bmp, image/tiff, image/x-portable-pixmap

Bytes per image: Unknown

Aggregate encoded bytes: Unknown

Hard width / height: Unknown / Unknown

Aggregate base64 image data must be less than 10 MB; MB definition unspecified. URL images less than 5 MB and download within 1.5 seconds.

No numeric universal provider resize recommendation. Pumpkin uses its conservative resize policy.

Pumpkin-selected image limits, not provider maxima:

• Pixel area: 2500000
• Decoded allocation: 100000000 bytes
• Aggregate encoded image data: 9000000 bytes
• Output format: image/png

Preserve aspect ratio; never upscale. Static GIF/WebP can be normalized; animated GIF is rejected. Provider-listed BMP/TIFF/PPM does not imply Pumpkin decoder support. Unsupported inputs fail explicitly.

Editorial guidance: DeepSeek V4 Flash with experimental image input; cheap, 1M context.

Provider output limit Unknown. thinkingMode=none and the 16384 output cap describe Pumpkin adapter policy, not absence of provider reasoning.

Exact context Unknown until metadata discovery; rounded public context labels are not treated as exact limits.

Serverless retirement announced for 2026-09-25; exact cutoff time unknown. Dedicated deployments are excluded.

Verified Sep 22, 2026

Provider output: Unknown separate numeric maximum

Streaming: Unknown

Effort values: Unknown

Editorial guidance: OpenAI open-weight 120B: solid reasoning and tool use at low cost; no images.

Provider output limit Unknown. thinkingMode=none and the 16384 output cap describe Pumpkin adapter policy, not absence of provider reasoning.

Exact context Unknown until metadata discovery; rounded public context labels are not treated as exact limits.

Verified Sep 22, 2026

Provider output: Unknown separate numeric maximum

Streaming: Unknown

Effort values: Unknown

Provider formats: image/png, image/jpeg, image/gif, image/bmp, image/tiff, image/x-portable-pixmap

Bytes per image: Unknown

Aggregate encoded bytes: Unknown

Hard width / height: Unknown / Unknown

Aggregate base64 image data must be less than 10 MB; MB definition unspecified. URL images less than 5 MB and download within 1.5 seconds.

No numeric universal provider resize recommendation. Pumpkin uses its conservative resize policy.

Pumpkin-selected image limits, not provider maxima:

• Pixel area: 2500000
• Decoded allocation: 100000000 bytes
• Aggregate encoded image data: 9000000 bytes
• Output format: image/png

Preserve aspect ratio; never upscale. Static GIF/WebP can be normalized; animated GIF is rejected. Provider-listed BMP/TIFF/PPM does not imply Pumpkin decoder support. Unsupported inputs fail explicitly.

Editorial guidance: Small, cheap GLM with image input: quick edits and lookups.

Provider output limit Unknown. thinkingMode=none and the 16384 output cap describe Pumpkin adapter policy, not absence of provider reasoning.

Exact context Unknown until metadata discovery; rounded public context labels are not treated as exact limits.

Verified Sep 22, 2026

Provider output: Unknown separate numeric maximum

Streaming: Unknown

Effort values: Unknown

Provider formats: image/png, image/jpeg, image/gif, image/bmp, image/tiff, image/x-portable-pixmap

Bytes per image: Unknown

Aggregate encoded bytes: Unknown

Hard width / height: Unknown / Unknown

Aggregate base64 image data must be less than 10 MB; MB definition unspecified. URL images less than 5 MB and download within 1.5 seconds.

No numeric universal provider resize recommendation. Pumpkin uses its conservative resize policy.

Pumpkin-selected image limits, not provider maxima:

• Pixel area: 2500000
• Decoded allocation: 100000000 bytes
• Aggregate encoded image data: 9000000 bytes
• Output format: image/png

Preserve aspect ratio; never upscale. Static GIF/WebP can be normalized; animated GIF is rejected. Provider-listed BMP/TIFF/PPM does not imply Pumpkin decoder support. Unsupported inputs fail explicitly.

Editorial guidance: Small 30B always-on agent model with image input: cheap and quick for simple tasks.

Provider output limit Unknown. thinkingMode=none and the 16384 output cap describe Pumpkin adapter policy, not absence of provider reasoning.

Exact context Unknown until metadata discovery; rounded public context labels are not treated as exact limits.

Serverless retirement announced for 2026-09-25; exact cutoff time unknown. Dedicated deployments are excluded.

Verified Sep 22, 2026

Provider output: Unknown separate numeric maximum

Streaming: Unknown

Effort values: Unknown

Editorial guidance: Cheapest fast agent model (30B, 3B active): small edits, descriptions, triage.

Provider output limit Unknown. thinkingMode=none and the 16384 output cap describe Pumpkin adapter policy, not absence of provider reasoning.

Exact context Unknown until metadata discovery; rounded public context labels are not treated as exact limits.

Verified Sep 22, 2026

Provider output: Unknown separate numeric maximum

Streaming: Unknown

Effort values: Unknown

Provider formats: image/png, image/jpeg, image/gif, image/bmp, image/tiff, image/x-portable-pixmap

Bytes per image: Unknown

Aggregate encoded bytes: Unknown

Hard width / height: Unknown / Unknown

Aggregate base64 image data must be less than 10 MB; MB definition unspecified. URL images less than 5 MB and download within 1.5 seconds.

No numeric universal provider resize recommendation. Pumpkin uses its conservative resize policy.

Pumpkin-selected image limits, not provider maxima:

• Pixel area: 2500000
• Decoded allocation: 100000000 bytes
• Aggregate encoded image data: 9000000 bytes
• Output format: image/png

Preserve aspect ratio; never upscale. Static GIF/WebP can be normalized; animated GIF is rejected. Provider-listed BMP/TIFF/PPM does not imply Pumpkin decoder support. Unsupported inputs fail explicitly.

Editorial guidance: Fireworks serverless model with image input and a 1M context; capability not rated.

Provider output limit Unknown. thinkingMode=none and the 16384 output cap describe Pumpkin adapter policy, not absence of provider reasoning.

Exact context Unknown until metadata discovery; rounded public context labels are not treated as exact limits.

Verified Sep 22, 2026

Provider output: 128000

Streaming: Yes

Effort values: none, low, medium, high, xhigh, max

Provider formats: image/png, image/jpeg, image/webp, image/gif

Bytes per image: Unknown

Aggregate encoded bytes: Unknown

Hard width / height: Unknown / Unknown

GIF must be nonanimated. The image guide allows up to 512 MB total request payload; exact byte accounting is unspecified. Model-specific image dimensions remain Unknown.

Pumpkin-selected image limits, not provider maxima:

• Pixel area: 2500000
• Decoded allocation: 100000000 bytes
• Aggregate encoded image data: 20000000 bytes
• Output format: image/png

Preserve aspect ratio; never upscale. Static GIF/WebP can be normalized; animated GIF is rejected. Provider-listed BMP/TIFF/PPM does not imply Pumpkin decoder support. Unsupported inputs fail explicitly.

Editorial guidance: Complex coding and agent work when Astra is more than you need. Cheaper than GPT-5.6 Sol.

Use Responses for built-in tools and function calling. Chat Completions function calling requires reasoning_effort=none.

Exact model-specific image resize limits are not stated in the reviewed image guide; Pumpkin uses its conservative client policy, not an inferred provider maximum.

Verified Sep 22, 2026

Provider output: 128000

Streaming: Yes

Effort values: none, low, medium, high, xhigh, max

Provider formats: image/png, image/jpeg, image/webp, image/gif

Bytes per image: Unknown

Aggregate encoded bytes: Unknown

Hard width / height: Unknown / Unknown

GIF must be nonanimated. The image guide allows up to 512 MB total request payload; exact byte accounting is unspecified. Model-specific image dimensions remain Unknown.

Pumpkin-selected image limits, not provider maxima:

• Pixel area: 2500000
• Decoded allocation: 100000000 bytes
• Aggregate encoded image data: 20000000 bytes
• Output format: image/png

Preserve aspect ratio; never upscale. Static GIF/WebP can be normalized; animated GIF is rejected. Provider-listed BMP/TIFF/PPM does not imply Pumpkin decoder support. Unsupported inputs fail explicitly.

Editorial guidance: Cheapest OpenAI: high-volume tasks, classification, short rewrites. Not for design or long hunts.

Use Responses for built-in tools and function calling. Chat Completions function calling requires reasoning_effort=none.

Exact model-specific image resize limits are not stated in the reviewed image guide; Pumpkin uses its conservative client policy, not an inferred provider maximum.

Verified Sep 22, 2026