ZMIME
Comparison · 3 models · Updated Oct 4, 2026

Claude Haiku 3.5 vs Gemma 3 12B IT vs Llama-3.3-70B-Instruct

Claude Haiku 3.5 comes out ahead, 49 to 43 and 41 on our weighted score.

  1. Our pick

    Anthropic

    Claude Haiku 3.5

    Released Oct 22, 2024

    49/100
    • ECI127.2
    • Price—
    • Context200K
    Full specs Remove Claude Haiku 3.5

  2. Google

    Gemma 3 12B IT

    Released Mar 12, 2025

    43/100
    • ECI123.5
    • Price$0.05 / $0.15
    • Context131K
    Full specs Remove Gemma 3 12B IT

  3. Meta

    Llama-3.3-70B-Instruct

    Released Dec 6, 2024

    41/100
    • ECI127.3
    • Price$0.59 / $0.724
    • Context128K
    Full specs Remove Llama-3.3-70B-Instruct
01 — Verdict

Claude Haiku 3.5 is our pick

Claude Haiku 3.5 is the better all-round choice, scoring 49/100 against Gemma 3 12B IT (43) and Llama-3.3-70B-Instruct (41). It leads on inputs & features and context window. The score weighs capability 67%, inputs & features 20%, context window 13%.

  • CapabilityLlama-3.3-70B-InstructCapabilities Index (ECI): Llama-3.3-70B-Instruct 127.3 · Claude Haiku 3.5 127.2 · Gemma 3 12B IT 123.5
  • Lowest priceGemma 3 12B ITGemma 3 12B IT $0.075 · Llama-3.3-70B-Instruct $0.624 per 1M tokens (3:1 blend) · Claude Haiku 3.5 unpriced
  • Longest contextClaude Haiku 3.5Claude Haiku 3.5 200,000 · Gemma 3 12B IT 131,072 · Llama-3.3-70B-Instruct 128,000 tokens
  • Widest inputsClaude Haiku 3.5Claude Haiku 3.5: Text, Images, PDFs · Gemma 3 12B IT: Text, Images · Llama-3.3-70B-Instruct: Text
  • Self-hostingGemma 3 12B IT and Llama-3.3-70B-InstructPublishes downloadable weights
How the score is built
MeasureWeightClaude Haiku 3.5Gemma 3 12B ITLlama-3.3-70B-Instruct
CapabilityCapabilities Index (ECI)67%494549
Inputs & features20%605025
Context window13%322424
Overall100%49/10043/10041/100

Left out because at least one model lacks the data: price. The remaining weights were rescaled.

02 — Side by side

Every spec in one table

Highlighted cells lead their row. Dashes mean the data is not published.

Claude Haiku 3.5 vs Gemma 3 12B IT vs Llama-3.3-70B-Instruct specifications side by side
SpecificationClaude Haiku 3.5AnthropicGemma 3 12B ITGoogleLlama-3.3-70B-InstructMeta
Capability
Capabilities Index (ECI)127.2123.5127.3 (best)
ECI rank#134 of 148#138 of 148#133 of 148 (best)
GPQA DiamondGraduate-level science questions38.1%39.5%47.4% (best)
OTIS Mock AIME 2024–2025Competition mathematics4.3%16.7% (best)5.1%
Price per million tokens
Input—$0.05 (best)$0.59
Output—$0.15 (best)$0.724
Cached input———
Blended (3:1)—$0.075 (best)$0.624
Long-context rate—Same rateSame rate
Price source—Median of 7 providersMedian of 21 providers
Limits
Context window200,000 tokens (best)131,072 tokens128,000 tokens
Max output8,192 tokens131,072 tokens (best)4,096 tokens
Inputs and features
TextYesYesYes
ImagesYesYesNo
PDFsYesNoNo
AudioNoNoNo
VideoNoNoNo
ReasoningNoNoNo
Tool callingYesYesYes
Structured outputNoNoNo
Availability
WeightsProprietaryOpenOpen
API model ID——llama-3.3-70b-instruct
API providers—724 (best)
ReleasedOct 22, 2024Mar 12, 2025Dec 6, 2024
Knowledge cutoffJul 31, 2024Aug 2024Dec 2023
03 — Cost

What would a month cost?

Enter your expected volume in millions of tokens. List prices only; caching and batch discounts would lower these.

  • Claude Haiku 3.5—
  • Gemma 3 12B IT$0.80
  • Llama-3.3-70B-Instruct$7.35
04 — Questions

Which should you choose?

Which is better: Claude Haiku 3.5, Gemma 3 12B IT or Llama-3.3-70B-Instruct?

Claude Haiku 3.5 is the better all-round choice, scoring 49/100 against Gemma 3 12B IT (43) and Llama-3.3-70B-Instruct (41). It leads on inputs & features and context window. The score weighs capability 67%, inputs & features 20%, context window 13%.

Which is cheaper, Claude Haiku 3.5, Gemma 3 12B IT or Llama-3.3-70B-Instruct?

Gemma 3 12B IT is cheaper at $0.05 input / $0.15 output per million tokens (median across 7 API providers). Llama-3.3-70B-Instruct costs $0.59 input / $0.724 output per million tokens (median across 21 API providers; free on Llama). At a typical mix of three input tokens to one output token, that is $0.075 per million tokens for Gemma 3 12B IT versus $0.624 for Llama-3.3-70B-Instruct (8.3× as much). Claude Haiku 3.5 has no published per-token price.

Which scores higher on benchmarks?

Llama-3.3-70B-Instruct scores higher on the Capabilities Index (ECI): Llama-3.3-70B-Instruct 127.3 (#133 of 148), Claude Haiku 3.5 127.2 (#134 of 148) and Gemma 3 12B IT 123.5 (#138 of 148). The confidence ranges of the top two overlap (122.5–129.5 vs 120.7–129.7), so treat the gap as small. On individual benchmarks: GPQA Diamond — Llama-3.3-70B-Instruct 47.4%, Gemma 3 12B IT 39.5%, Claude Haiku 3.5 38.1%; OTIS Mock AIME 2024–2025 — Gemma 3 12B IT 16.7%, Llama-3.3-70B-Instruct 5.1%, Claude Haiku 3.5 4.3%.

Which is better for coding?

There are no published SWE-bench Verified results for Claude Haiku 3.5, Gemma 3 12B IT and Llama-3.3-70B-Instruct yet, so there is no like-for-like coding score. On overall capability, Llama-3.3-70B-Instruct 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?

Claude Haiku 3.5 has the largest context window at 200,000 tokens, against 131,072 for Gemma 3 12B IT and 128,000 for Llama-3.3-70B-Instruct. Maximum output per response: Claude Haiku 3.5 up to 8,192, Gemma 3 12B IT up to 131,072, Llama-3.3-70B-Instruct up to 4,096 tokens.

Which can read images, PDFs, audio or video?

Claude Haiku 3.5 accepts text, images and PDFs; Gemma 3 12B IT accepts text and images; Llama-3.3-70B-Instruct accepts text. Claude Haiku 3.5 handles the widest range of inputs.

Are any of these open source?

Gemma 3 12B IT and Llama-3.3-70B-Instruct publishes its weights and can be self-hosted; Claude Haiku 3.5 is proprietary.

Which is newer?

Gemma 3 12B IT is the newest, released Mar 12, 2025. Llama-3.3-70B-Instruct came out Dec 6, 2024; Claude Haiku 3.5 came out Oct 22, 2024. Knowledge cutoff: Claude Haiku 3.5 Jul 31, 2024, Gemma 3 12B IT Aug 2024, Llama-3.3-70B-Instruct Dec 2023.

How do you decide the winner?

Each model gets a 0–100 score on 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 Oct 4, 2026.

05 — Keep comparing

Related comparisons