DeepSeek V4.1 Flash 1017 vs Gemini 4 Carbon vs Claude Haiku 5.5
Claude Haiku 5.5 comes out ahead, 82 to 79 and 75 on our weighted score, and it is the cheaper option too.Gemini 4 Carbon comes out ahead, 93 to 83 and 80 on our weighted score, though Claude Haiku 5.5 is 20× cheaper per token.
DeepSeek
DeepSeek V4.1 Flash 1017
79/100- ECI159.4
- Price$0.15 / $0.60
- Context1M
Google
Gemini 4 Carbon
75/100- ECI169.7
- Price$2.00 / $10.00
- Context1.05M
- Our pick
Anthropic
Claude Haiku 5.5
82/100- ECI160.3
- Price$0.10 / $0.50
- Context1M
DeepSeek
DeepSeek V4.1 Flash 1017
80/100- ECI159.4
- Price$0.15 / $0.60
- Context1M
- Our pick
Google
Gemini 4 Carbon
93/100- ECI169.7
- Price$2.00 / $10.00
- Context1.05M
Anthropic
Claude Haiku 5.5
83/100- ECI160.3
- Price$0.10 / $0.50
- Context1M
Claude Haiku 5.5 is our pickGemini 4 Carbon is our pick
Claude Haiku 5.5 is the better all-round choice, scoring 82/100 against DeepSeek V4.1 Flash 1017 (79) and Gemini 4 Carbon (75). It leads on price. Gemini 4 Carbon wins on capability and inputs & features. The score weighs capability 50%, price 25%, inputs & features 15%, context window 10%.Gemini 4 Carbon is the better all-round choice, scoring 93/100 against Claude Haiku 5.5 (83) and DeepSeek V4.1 Flash 1017 (80). It leads on capability and inputs & features. The score weighs capability 67%, inputs & features 20%, context window 13%.
- CapabilityGemini 4 CarbonCapabilities Index (ECI): Gemini 4 Carbon 169.7 · Claude Haiku 5.5 160.3 · DeepSeek V4.1 Flash 1017 159.4
- Lowest priceClaude Haiku 5.5Claude Haiku 5.5 $0.20 · DeepSeek V4.1 Flash 1017 $0.263 · Gemini 4 Carbon $4.00 per 1M tokens (3:1 blend)
- Longest contextGemini 4 CarbonGemini 4 Carbon 1,048,576 · DeepSeek V4.1 Flash 1017 1,000,000 · Claude Haiku 5.5 1,000,000 tokens
- Widest inputsGemini 4 CarbonDeepSeek V4.1 Flash 1017: Text, Images · Gemini 4 Carbon: Text, Images, PDFs, Audio, Video · Claude Haiku 5.5: Text, Images, PDFs
- Self-hostingDeepSeek V4.1 Flash 1017Publishes downloadable weights (MIT)
| Measure | Weight | DeepSeek V4.1 Flash 1017 | Gemini 4 Carbon | Claude Haiku 5.5 |
|---|---|---|---|---|
| CapabilityCapabilities Index (ECI) | 50% | 87 | 97 | 88 |
| Price | 25% | 74 | 20 | 79 |
| Inputs & features | 15% | 70 | 100 | 80 |
| Context window | 10% | 64 | 65 | 64 |
| Overall | 100% | 79/100 | 75/100 | 82/100 |
| Measure | Weight | DeepSeek V4.1 Flash 1017 | Gemini 4 Carbon | Claude Haiku 5.5 |
|---|---|---|---|---|
| CapabilityCapabilities Index (ECI) | 67% | 87 | 97 | 88 |
| Inputs & features | 20% | 70 | 100 | 80 |
| Context window | 13% | 64 | 65 | 64 |
| Overall | 100% | 80/100 | 93/100 | 83/100 |
Every spec in one table
Highlighted cells lead their row. Unavailable specifications are marked “Not published”. Publisher results include their evaluation settings and source.
| Specification | |||
|---|---|---|---|
| Capability | |||
| Capabilities Index (ECI) | 159.4 | 169.7 (best) | 160.3 |
| ECI rank | #24 of 159 | #3 of 159 (best) | #20 of 159 |
| ECI confidence interval | 157.1–161.9 | 166.6–173.1 | 158.2–162.7 |
| ECI evaluation date | Oct 17, 2026 | Nov 10, 2026 | Oct 7, 2026 |
| Documented capabilities | October 17 update of DeepSeek V4.1 Flash with improved reasoning, coding, agentic tasks, and visual understanding | Google's most capable reasoning model for complex software engineering, deep math and science, high factual accuracy, and long-horizon autonomous agents | Fast Claude model for responsive assistance, classification, and lightweight agents |
| Model parameters | Not published | Not published | Not published |
| Output modalities | Text | Text | Text |
| GPQA DiamondGraduate-level science questions | 93.2%max | 96.5%high (best) | 89.7%max |
| FrontierMath Tiers 1–3Research-level mathematics | 71.2%max | 90.5%high (best) | 75.1%max |
| OTIS Mock AIME 2024–2025Competition mathematics | 99.4%max | 100%high (best) | 98.9%xhigh |
| SimpleQA VerifiedShort factual questions | 42.5%max | 77.7%high (best) | 23.8%max |
| Publisher-reported benchmarks | |||
| Humanity's Last Exam · full setwithout tools | 42.6Reported result | Not published | Not published |
| Humanity's Last Exam · text-only subsetwithout tools | 44Reported result | Not published | Not published |
| Terminal-Bench 4.0Publisher-reported result | 42.3Reported result | Not published | Not published |
| DeepSWE 1.1Publisher-reported result | 77Reported result | Not published | Not published |
| ProgramBenchPublisher-reported result | 29.1Reported result | Not published | Not published |
| NL2RepoPublisher-reported result | 68.2Reported result | Not published | Not published |
| AutomationBench 1.0.6Publisher-reported result | 60.5Reported result | Not published | Not published |
| Agents' Last ExamPublisher-reported result | 35.4Reported result | Not published | Not published |
| Apex Bench 1.1Publisher-reported result | 72.1Reported result | Not published | Not published |
| Chartographywith tools | 83Reported result | Not published | Not published |
| BabyVisionwith tools | 92.5Reported result | Not published | Not published |
| ZeroBench · mainwith tools | 51pass@5Reported result | Not published | Not published |
| GDPval-AA 2.1Publisher-reported result | Not published | Not published | 1620EloReported resultOct 7, 2026 |
| AA-Briefcase 1.1Publisher-reported result | Not published | Not published | 1578EloReported resultOct 7, 2026 |
| OSWorld 2.1offline subset | Not published | Not published | 72.4partial scoreReported resultOct 7, 2026 |
| Humanity's Last Examno tools | Not published | Not published | 45.9scoreReported resultOct 7, 2026 |
| Humanity's Last Examwith tools | Not published | Not published | 57.4scoreReported resultOct 7, 2026 |
| Terminal-Bench 4.0Publisher-reported result | Not published | Not published | 39.2scoreReported resultOct 7, 2026 |
| FrontierCode 1.1 · MainPublisher-reported result | Not published | Not published | 46.4scoreReported resultOct 7, 2026 |
| Chartographyno tools | Not published | Not published | 46.4scoreReported resultOct 7, 2026 |
| Price per million tokens | |||
| Input | $0.15 | $2.00 | $0.10 (best) |
| Output | $0.60 | $10.00 | $0.50 (best) |
| Cached input | $0.003 (best) | $0.10 | $0.01 |
| Blended (3:1) | $0.263 | $4.00 | $0.20 (best) |
| Long-context rate | Same rate | Same rate | Over 100K: $0.50 / $2.50 |
| Price source | Official DeepSeek API | Official Google API | Official Anthropic API |
| Limits | |||
| Context window | 1,000,000 tokens | 1,048,576 tokens (best) | 1,000,000 tokens |
| Max input | Not published | Not published | Not published |
| Max output | 384,000 tokens (best) | 65,536 tokens | 128,000 tokens |
| Inputs and features | |||
| Text | Yes | Yes | Yes |
| Images | Yes | Yes | Yes |
| PDFs | No | Yes | Yes |
| Audio | No | Yes | No |
| Video | No | Yes | No |
| Reasoning | Yeslow · high · max | Yeslow · medium · high | Yeslow · medium · high · xhigh · max |
| Tool calling | Yes | Yes | Yes |
| Structured output | Yes | Yes | Yes |
| File attachments | Yes | Yes | Yes |
| Temperature control | Yes | Yes | No |
| Reasoning control | effort | effort | effort |
| Availability | |||
| Weights | OpenMIT | Proprietary | Proprietary |
| API model ID | deepseek-flash | gemini-4-carbon | claude-haiku-5-5 |
| API providers | 83 (best) | 19 | 48 |
| Released | Oct 17, 2026 | Nov 10, 2026 | Oct 7, 2026 |
| Knowledge cutoff | May 2025 | Mar 2026 | Jun 2026 |
| Publisher documentation | Publisher documentation · Model weights | Publisher documentation | Publisher documentation |
| Source metadata updated | Oct 17, 2026 | Nov 10, 2026 | Oct 7, 2026 |
| Data sources | models.dev · Epoch AI | models.dev · Epoch AI | models.dev · Epoch AI |
| Catalog checked | Nov 22, 2026 | Nov 22, 2026 | Nov 22, 2026 |
What would a month cost?
Enter your expected volume in millions of tokens. List prices only; caching and batch discounts would lower these.
DeepSeek V4.1 Flash 1017$2.70
Gemini 4 Carbon$40.00
Claude Haiku 5.5$2.00
Which should you choose?
Which is better: DeepSeek V4.1 Flash 1017, Gemini 4 Carbon or Claude Haiku 5.5?
Claude Haiku 5.5 is the better all-round choice, scoring 82/100 against DeepSeek V4.1 Flash 1017 (79) and Gemini 4 Carbon (75). It leads on price. Gemini 4 Carbon wins on capability and inputs & features. The score weighs capability 50%, price 25%, inputs & features 15%, context window 10%.Gemini 4 Carbon is the better all-round choice, scoring 93/100 against Claude Haiku 5.5 (83) and DeepSeek V4.1 Flash 1017 (80). It leads on capability and inputs & features. The score weighs capability 67%, inputs & features 20%, context window 13%.
Which is cheaper, DeepSeek V4.1 Flash 1017, Gemini 4 Carbon or Claude Haiku 5.5?
Claude Haiku 5.5 is cheaper at $0.10 input / $0.50 output per million tokens (official Anthropic API price). DeepSeek V4.1 Flash 1017 costs $0.15 input / $0.60 output per million tokens (official DeepSeek API price); Gemini 4 Carbon costs $2.00 input / $10.00 output per million tokens (official Google API price). At a typical mix of three input tokens to one output token, that is $0.20 per million tokens for Claude Haiku 5.5 versus $0.263 for DeepSeek V4.1 Flash 1017 (1.3× as much) and $4.00 for Gemini 4 Carbon (20× as much).
Which scores higher on benchmarks?
Gemini 4 Carbon scores higher on the Capabilities Index (ECI): Gemini 4 Carbon 169.7 (#3 of 159), Claude Haiku 5.5 160.3 (#20 of 159) and DeepSeek V4.1 Flash 1017 159.4 (#24 of 159). Their confidence ranges do not overlap (166.6–173.1 vs 158.2–162.7), so the gap is a real one. On individual benchmarks: GPQA Diamond — Gemini 4 Carbon 96.5%, DeepSeek V4.1 Flash 1017 93.2%, Claude Haiku 5.5 89.7%; FrontierMath Tiers 1–3 — Gemini 4 Carbon 90.5%, Claude Haiku 5.5 75.1%, DeepSeek V4.1 Flash 1017 71.2%; OTIS Mock AIME 2024–2025 — Gemini 4 Carbon 100%, DeepSeek V4.1 Flash 1017 99.4%, Claude Haiku 5.5 98.9%; SimpleQA Verified — Gemini 4 Carbon 77.7%, DeepSeek V4.1 Flash 1017 42.5%, Claude Haiku 5.5 23.8%.
Which is better for coding?
There are no published SWE-bench Verified results for DeepSeek V4.1 Flash 1017, Gemini 4 Carbon and Claude Haiku 5.5 yet, so there is no like-for-like coding score. On overall capability, Gemini 4 Carbon leads, which tends to carry over to coding, but test on your own codebase. All three support tool calling for agent workflows.
Which has the bigger context window?
Gemini 4 Carbon has the largest context window at 1,048,576 tokens, against 1,000,000 for DeepSeek V4.1 Flash 1017 and 1,000,000 for Claude Haiku 5.5. Maximum output per response: DeepSeek V4.1 Flash 1017 up to 384,000, Gemini 4 Carbon up to 65,536, Claude Haiku 5.5 up to 128,000 tokens.
Which can read images, PDFs, audio or video?
DeepSeek V4.1 Flash 1017 accepts text and images; Gemini 4 Carbon accepts text, images, PDFs, audio and video; Claude Haiku 5.5 accepts text, images and PDFs. Gemini 4 Carbon handles the widest range of inputs.
Are any of these open source?
DeepSeek V4.1 Flash 1017 publishes its weights (MIT) and can be self-hosted; Gemini 4 Carbon and Claude Haiku 5.5 is proprietary.
Which is newer?
Gemini 4 Carbon is the newest, released Nov 10, 2026. DeepSeek V4.1 Flash 1017 came out Oct 17, 2026; Claude Haiku 5.5 came out Oct 7, 2026. Knowledge cutoff: DeepSeek V4.1 Flash 1017 May 2025, Gemini 4 Carbon Mar 2026, Claude Haiku 5.5 Jun 2026.
How do you decide the winner?
This rating uses capability (50%, independent benchmark results); price (25%, blended price per million tokens (3 input : 1 output), log scale); inputs & features (15%, image, PDF, audio and video input, tool calling, structured output and reasoning); context window (10%, maximum tokens per request, log scale). Dimensions missing for any model are dropped and the remaining weights rescaled, so every model is judged on the same evidence. Specs and prices come from public model listings and the labs’ own API pages; capability scores come from independent benchmark runs. Data updated Nov 22, 2026.This rating uses capability (67%, independent benchmark results); inputs & features (20%, image, PDF, audio and video input, tool calling, structured output and reasoning); context window (13%, maximum tokens per request, log scale). Dimensions missing for any model are dropped and the remaining weights rescaled, so every model is judged on the same evidence. Specs and prices come from public model listings and the labs’ own API pages; capability scores come from independent benchmark runs. Data updated Nov 22, 2026.