DeepSeek currently exposes two API model IDs: deepseek-flash, which serves DeepSeek V4.1 Flash (released September 10, 2026, with native image input), and deepseek-v4-pro, which serves DeepSeek V4 Pro 0813. Both support a 1M-token context window and thinking and non-thinking modes. V4 Flash was retired on September 10; the legacy deepseek-v4-flash ID routes to V4.1 Flash. DeepSeek announced that deepseek-v4-pro requests would route there too from September 14, then said on its pricing page, read September 17, 2026, that V4 Pro service continues with billing unchanged. This directory tracks 13 variants across 11 model families.
The model name and the API ID are not always the same thing. DeepSeek uses one API ID per family, then switches thinking and reasoning effort through request settings. We keep those configurations as separate benchmark rows because their scores differ, but we do not pretend they are separate billable SKUs.
Older R1, V3.2, V3.1, Coder, and Math rows remain useful for reproducibility and self-hosting. They are not current first-party API choices. DeepSeek set July 24, 2026 as the retirement deadline for the deepseek-chat and deepseek-reasoner aliases, so new integrations should use the two IDs above.
Start with the family, then choose a reasoning setting. The public score is the strongest tracked configuration in each family, not a promise that every prompt or API mode will reproduce it.
| Family | API ID | Status | Public score | Context | Use it when |
|---|---|---|---|---|---|
| DeepSeek V4.1 Flash | deepseek-flash | Current | Not ranked | 1M | New work. DeepSeek's launch table puts it ahead of V4 Pro on every agentic coding, cyber, and tool-use row it reports, at a lower price, with image input. |
| DeepSeek V4 Pro | deepseek-v4-pro | Current | 65/100 | 1M | Existing integrations only. DeepSeek announced a September 14, 2026 reroute to V4.1 Flash, then kept V4 Pro in service with no end date published. |
| DeepSeek V4 Flash | deepseek-v4-flash | Retired | Not ranked | 1M | Nothing new. Retired September 10, 2026; its ID now routes to V4.1 Flash. |
Our default: start with V4.1 Flash and measure task failures against an acceptance test. Do not build new routing rules around V4 Pro: DeepSeek says V4.1 Flash beats it on performance, cost, speed, and total runtime. It cancelled the September 14 reroute it had announced, and its pricing page says only that it will give notice before V4 Pro service changes.
Honest limit. V4.1 Flash has no public score yet. Its launch rows are provider-run and harness-specific, so they stay display-only until benchmark-owned or independent evaluator results land. If latency or regional availability decides the deployment, run the same workload against the live endpoint before choosing.
| Release | Date | What changed | Current role |
|---|---|---|---|
| DeepSeek V4.1 Flash | Sep 10, 2026 | New Causal Encoder-Decoder architecture (552B MoE, 8B active for input and 16B for output), native image input, MIT weights, and the deepseek-flash ID. V4 Flash retired the same day. | Current API and open-weight family |
| DeepSeek V4 | Apr 24, 2026 | Pro and Flash introduced dedicated IDs and a 1M-token context window. | V4 Pro still served, September 14 reroute cancelled; V4 Flash retired |
| DeepSeek V3.2 | Dec 1, 2025 | Unified thinking and non-thinking modes in the V3 line. | Historical and self-hosted comparison |
| DeepSeek R1 | Jan 20, 2025 | Released the reasoning model and distilled open-weight variants. | Self-hosting and reasoning-history reference |
| DeepSeek V3 | Dec 26, 2024 | Established the 671B mixture-of-experts V3 family. | Historical and open-weight reference |
DeepSeek V4.1 Flash is the newest model, and DeepSeek's own launch table puts it ahead of V4 Pro on every agentic row it reports. V4 Pro 0813 still holds the stronger tracked public score at 65/100 because V4.1 Flash's launch rows are provider-run and stay display-only. Use V4.1 Flash for new work and escalate only the requests that fail your acceptance tests.
The current first-party API documentation lists deepseek-flash (DeepSeek V4.1 Flash) and deepseek-v4-pro (DeepSeek V4 Pro 0813). Each ID supports thinking and non-thinking modes. The legacy deepseek-v4-flash and deepseek-v4-flash-vision-exp IDs route to V4.1 Flash. DeepSeek announced that deepseek-v4-pro would join them from September 14, 2026, then stated on its pricing page that V4 Pro service continues. The older deepseek-chat and deepseek-reasoner aliases reached their retirement deadline on July 24, 2026.
V4.1 Flash is the first model in DeepSeek's new architecture family: a 552B mixture-of-experts model that activates 8B parameters per input token and 16B per output token, reads images natively, and costs less per token than V4 Pro. V4 Pro is the larger previous-generation model. Both expose a 1M-token context window and configurable thinking modes.
DeepSeek publishes weights for V4.1 Flash (MIT license), V4, R1, V3.2, and several earlier families. We label them open weight: downloadable weights are clear evidence, while full training-data and reproducibility openness is a broader claim. Hosted API use still incurs token charges.
Current provider directory
As of September 27, 2026, this directory contains 2 active, preview, announced, or limited-access entries. Provider-issued IDs appear only when the underlying documentation identifies them; BenchLM model keys are kept separate.
Sep 10, 2026
Open the dated sourceOperational record
Shared IDs are listed once. Request-time reasoning modes and other configurations do not become separate model IDs.
deepseek-flashSource for deepseek-flashdeepseek-v4-proSource for deepseek-v4-proTogether with the latest release above, these are the three newest dated release groups in the tracked provider inventory.
DeepSeek V4 Pro 0813
Aug 13, 2026
Dated sourceDeepSeek V4 Flash 0731
Jul 31, 2026
Dated sourceSearch every tracked model variant, including preview, limited-access, deprecated, and open-weight releases. Provider identifiers and specifications link to first-party evidence where it is available.
Showing 13 of 13 variants
Prices are input / output per 1M tokens.
deepseek-flashdeepseek-ai/DeepSeek-V4.1-FlashThinking mode at max effort on the deepseek-flash API ID. The model card evaluates at reasoning_effort=100; the API exposes low, high, and max effort levels, with max mapped from the highest requested settings.
text, image → text
input → output
deepseek-v4-proThinking mode with max effort on the shared family API ID.
text → text
input → output
Not documented
input → output
Not documented
input → output
Not documented
input → output
Not documented
input → output
Not documented
input → output
Not documented
input → output
Not documented
input → output
deepseek-v4-flashThinking mode with max effort on the shared family API ID.
text → text
input → output
Not documented
input → output
Not documented
input → output
Not documented
input → output
Prices are input / output per 1M tokens.
DeepSeek's average score across its top 3 ranked models, cumulative through each release month.
One canonical entry per model family. Public scores use the same evidence rules as the main leaderboard.
| Context | Price | Public score | |||||
|---|---|---|---|---|---|---|---|
| DeepSeek V4 Pro 0813 | Current | 1M | $1.32 / $3.96 | 66 | 31.79s | 63.48Estimated | |
| DeepSeek V3.2 | Established | 128K | $0.28 / $0.42 | 35 | 3.75s | 50.21Supported | |
| DeepSeek-R1 | Established | 128K | $0.55 / $2.19 | N/A | N/A | 42.63Supported | |
| DeepSeek V3.1 | Established | 128K | $0.00 / $0.00 | N/A | N/A | 41.27Supported | |
| DeepSeek V3 | Established | 128K | $0.27 / $1.10 | N/A | N/A | 31.9Supported | |
| DeepSeek R1 Distill Qwen 32B | Established | 128K | $0.00 / $0.00 | 60 | 0.84s | 30.45Estimated | |
| DeepSeek Coder 2.0 | Tracked | 128K | N/A | N/A | N/A | — | |
| DeepSeek LLM 2.0 | Tracked | 128K | $0.00 / $0.00 | N/A | N/A | — | |
| DeepSeek V4 Flash 0731 | Superseded | 1M | $0.14 / $0.28 | 221 | 10.17s | — | |
| DeepSeek V4.1 Flash | Current | 1M | $0.30 / $1.20 | 227 | 9.90s | — | |
| DeepSeekMath V2 | Tracked | 128K | $0.00 / $0.00 | N/A | N/A | — |
DeepSeek · Model release
DeepSeek · Model release
DeepSeek · Model release
DeepSeek · Model release
DeepSeek · Model release
Radar confirmed these at the source. Use DeepSeek in your work? Explore Radar to follow supported changes and choose your alerts.